TF: T5 can now handle a padded past (i.e. XLA generation) (#17969)

* get the right slicing index for position_bias

src/transformers/models/t5/modeling_tf_t5.py

@@ -23,6 +23,7 @@ from typing import Optional, Tuple, Union
 
 import numpy as np
 import tensorflow as tf
+from tensorflow.compiler.tf2xla.python.xla import dynamic_slice
 
 from ...activations_tf import get_tf_activation
 from ...modeling_tf_outputs import (
@@ -384,10 +385,19 @@ class TFT5Attention(tf.keras.layers.Layer):
             else:
                 position_bias = self.compute_bias(real_seq_length, key_length)
 
-            # if key and values are already calculated
-            # we want only the last query position bias
+            # if key and values are already calculated we want only the last query position bias
             if past_key_value is not None:
+                if not self.has_relative_attention_bias:
                     position_bias = position_bias[:, :, -seq_length:, :]
+                else:
+                    # we might have a padded past structure, in which case we want to fetch the position bias slice
+                    # right after the most recently filled past index
+                    most_recently_filled_past_index = tf.reduce_max(tf.where(past_key_value[0][0, 0, :, 0] != 0.0))
+                    position_bias = dynamic_slice(
+                        position_bias,
+                        (0, 0, most_recently_filled_past_index + 1, 0),
+                        (1, self.n_heads, seq_length, real_seq_length),
+                    )
 
             if mask is not None:
                 position_bias = tf.cast(position_bias, dtype=mask.dtype)

tests/models/t5/test_modeling_tf_t5.py

@@ -590,21 +590,17 @@ class TFT5GenerationIntegrationTests(unittest.TestCase):
         ]
         input_ids = tokenizer(sentences, return_tensors="tf", padding=True).input_ids
 
-        # xla_generate = tf.function(model.generate, jit_compile=True)
-        xla_generate = tf.function(model.generate)
+        xla_generate = tf.function(model.generate, jit_compile=True)
 
-        # TODO (joao): there is something not quite right with XLA T5 -- as we increase `max_length` the two outputs
-        # drift appart, where the XLA version clearly degrades its quality. XLA-related variables look fine (they are
-        # being padded and filled in the right places). This also happens in other generation modes. Investigate.
-        output_ids = model.generate(input_ids, num_beams=2, max_length=9)
-        output_ids_xla = xla_generate(input_ids, num_beams=2, max_length=9)
+        output_ids = model.generate(input_ids, num_beams=2)
+        output_ids_xla = xla_generate(input_ids, num_beams=2)
 
         output_strings = tokenizer.batch_decode(output_ids, skip_special_tokens=True)
         output_strings_xla = tokenizer.batch_decode(output_ids_xla, skip_special_tokens=True)
 
         expected_output_string = [
             "Aujourd'hui est une belle journée.",
-            "J'ai quatre chats,",
+            "J'ai quatre chats, trois chiens, deux oiseaux et un cheval.",
         ]
 
         self.assertListEqual(expected_output_string, output_strings)