TF: XLA repetition penalty (#16879)

src/transformers/generation_tf_logits_process.py

@@ -241,18 +241,29 @@ class TFRepetitionPenaltyLogitsProcessor(TFLogitsProcessor):
 
         self.penalty = penalty
 
-    def _create_score_penalties(self, input_ids, logits):
+    def _create_score_penalties(self, input_ids: tf.Tensor, logits: tf.Tensor) -> tf.Tensor:
-        # create logit penalties for already seen input_ids
+        # We want to populate the penalties in the positions of `input_ids`. Since XLA can't handle shapes unknown
-        token_penalties = np.ones(logits.shape)
+        # before runtime, `tf.unique` can't be used. Therefore, we may have redundant updates, when a given row has
-        prev_input_ids = [np.unique(input_id) for input_id in input_ids.numpy()]
+        # the same token multiple times.
-        for i, prev_input_id in enumerate(prev_input_ids):
+
-            logit_penalized = logits[i].numpy()[prev_input_id]
+        # Gathers the penalties to apply
-            logit_penalties = np.zeros(logit_penalized.shape)
+        logit_penalties = tf.gather(logits, input_ids, axis=1, batch_dims=1)
-            # if previous logit score is < 0 then multiply repetition penalty else divide
+        logit_penalties = tf.where(logit_penalties > 0, 1 / self.penalty, logit_penalties)
-            logit_penalties[logit_penalized < 0] = self.penalty
+        logit_penalties = tf.where(logit_penalties < 0, self.penalty, logit_penalties)
-            logit_penalties[logit_penalized > 0] = 1 / self.penalty
+
-            np.put(token_penalties[i], prev_input_id, logit_penalties)
+        # Scatters the penalties
-        return tf.convert_to_tensor(token_penalties, dtype=tf.float32)
+        token_penalties = tf.ones(logits.shape)
+        indexable_prev_input_ids = tf.concat(
+            (
+                tf.expand_dims(tf.repeat(tf.range(input_ids.shape[0]), input_ids.shape[1]), axis=-1),
+                tf.expand_dims(tf.reshape(input_ids, [-1]), axis=-1),
+            ),
+            axis=1,
+        )
+        token_penalties = tf.tensor_scatter_nd_update(
+            token_penalties, indices=indexable_prev_input_ids, updates=tf.reshape(logit_penalties, [-1])
+        )
+        return token_penalties
 
     def __call__(self, input_ids: tf.Tensor, scores: tf.Tensor, cur_len: int) -> tf.Tensor:
         score_penalties = self._create_score_penalties(input_ids[:, :cur_len], scores)

tests/generation/test_generation_tf_logits_process.py

@@ -101,7 +101,7 @@ class TFLogitsProcessorTest(unittest.TestCase):
         self.assertGreater(tf.math.reduce_max(probs[1, :]), tf.math.reduce_max(warped_prob_smooth[1, :]))
         self.assertLess(tf.math.reduce_min(probs[1, :]), tf.math.reduce_min(warped_prob_smooth[1, :]))
 
-    def test_repetition_penalty_dist_process(self):
+    def _get_repetition_penalty_inputs(self):
         vocab_size = 10
         cur_len = 2
 
@@ -114,17 +114,31 @@ class TFLogitsProcessorTest(unittest.TestCase):
         scores = tf.where(mask, -1 / vocab_size, scores)
         mask = tf.cast(tf.constant([10 * [0], 5 * [0] + [1] + 4 * [0]]), tf.bool)
         scores = tf.where(mask, 4 / vocab_size, scores)
+        return vocab_size, cur_len, input_ids, scores
 
-        rep_penalty_proc = TFRepetitionPenaltyLogitsProcessor(penalty=2.0)
+    def _check_repetition_penalty_outputs(self, scores, vocab_size):
-
+        # check that values were correctly changed (negative scores for used tokens should increase, others
-        scores = rep_penalty_proc(input_ids, tf.identity(scores), cur_len)
+        # should decrease)
-
-        # check that values were correctly changed
         self.assertAlmostEqual(scores[0, 0].numpy(), -(1 / vocab_size) * 2)
         self.assertAlmostEqual(scores[0, 1].numpy(), (1 / vocab_size) / 2)
+        self.assertAlmostEqual(scores[0, 2].numpy(), (1 / vocab_size))  # unused tokens should see no change
 
         self.assertAlmostEqual(scores[1, 0].numpy(), (1 / vocab_size) / 2)
         self.assertAlmostEqual(scores[1, 5].numpy(), (4 / vocab_size) / 2)
+        self.assertAlmostEqual(scores[0, 2].numpy(), (1 / vocab_size))  # unused tokens should see no change
+
+    def test_repetition_penalty_dist_process(self):
+        vocab_size, cur_len, input_ids, scores = self._get_repetition_penalty_inputs()
+        rep_penalty_proc = TFRepetitionPenaltyLogitsProcessor(penalty=2.0)
+        scores = rep_penalty_proc(input_ids, tf.identity(scores), cur_len)
+        self._check_repetition_penalty_outputs(scores, vocab_size)
+
+    def test_repetition_penalty_dist_process_xla(self):
+        vocab_size, cur_len, input_ids, scores = self._get_repetition_penalty_inputs()
+        rep_penalty_proc = TFRepetitionPenaltyLogitsProcessor(penalty=2.0)
+        rep_penalty_proc = tf.function(rep_penalty_proc, jit_compile=True)  # added line wrt non-XLA test
+        scores = rep_penalty_proc(input_ids, tf.identity(scores), cur_len)
+        self._check_repetition_penalty_outputs(scores, vocab_size)
 
     def test_top_k_dist_warper(self):
         input_ids = None