SUMIN/HuggingFace_transformer
1
0
Fork 0
Code Issues Pull Requests Actions 7 Packages Projects Releases Wiki Activity

TF generate refactor - past without encoder outputs (#15944)

Browse Source 
* Remove packed past from generation_tf_utils

* update models with the new past format

* update template accordingly
This commit is contained in:
Joao Gante
2022-03-08 14:46:44 +00:00
committed by GitHub
parent 62d847602a
commit 70203b5937
30 changed files with 301 additions and 684 deletions
Download Patch File Download Diff File Expand all files Collapse all files

1
tests/bart/test_modeling_tf_bart.py
View File

@@ -116,7 +116,6 @@ class TFBartModelTester:
outputs = model(input_ids, attention_mask=attention_mask, head_mask=head_mask, use_cache=True)
output, past_key_values = outputs.to_tuple()
past_key_values = past_key_values[1]
# create hypothetical next token and extent to next_input_ids
next_tokens = ids_tensor((self.batch_size, 3), config.vocab_size)

1
tests/blenderbot/test_modeling_tf_blenderbot.py
View File

@@ -114,7 +114,6 @@ class TFBlenderbotModelTester:
outputs = model(input_ids, attention_mask=attention_mask, head_mask=head_mask, use_cache=True)
output, past_key_values = outputs.to_tuple()
past_key_values = past_key_values[1]
# create hypothetical next token and extent to next_input_ids
next_tokens = ids_tensor((self.batch_size, 3), config.vocab_size)

1
tests/blenderbot_small/test_modeling_tf_blenderbot_small.py
View File

@@ -114,7 +114,6 @@ class TFBlenderbotSmallModelTester:
outputs = model(input_ids, attention_mask=attention_mask, head_mask=head_mask, use_cache=True)
output, past_key_values = outputs.to_tuple()
past_key_values = past_key_values[1]
# create hypothetical next token and extent to next_input_ids
next_tokens = ids_tensor((self.batch_size, 3), config.vocab_size)

1
tests/led/test_modeling_tf_led.py
View File

@@ -133,7 +133,6 @@ class TFLEDModelTester:
outputs = model(input_ids, attention_mask=attention_mask, use_cache=True)
output, past_key_values = outputs.to_tuple()
past_key_values = past_key_values[1]
# create hypothetical next token and extent to next_input_ids
next_tokens = ids_tensor((self.batch_size, 3), config.vocab_size)

1
tests/marian/test_modeling_tf_marian.py
View File

@@ -116,7 +116,6 @@ class TFMarianModelTester:
outputs = model(input_ids, attention_mask=attention_mask, head_mask=head_mask, use_cache=True)
output, past_key_values = outputs.to_tuple()
past_key_values = past_key_values[1]
# create hypothetical next token and extent to next_input_ids
next_tokens = ids_tensor((self.batch_size, 3), config.vocab_size)

1
tests/pegasus/test_modeling_tf_pegasus.py
View File

@@ -114,7 +114,6 @@ class TFPegasusModelTester:
outputs = model(input_ids, attention_mask=attention_mask, head_mask=head_mask, use_cache=True)
output, past_key_values = outputs.to_tuple()
past_key_values = past_key_values[1]
# create hypothetical next token and extent to next_input_ids
next_tokens = ids_tensor((self.batch_size, 3), config.vocab_size)

2
tests/speech_to_text/test_modeling_tf_speech_to_text.py
View File

@@ -182,7 +182,7 @@ class TFSpeech2TextModelTester:
# first forward pass
outputs = model(input_ids, attention_mask=attention_mask, use_cache=True)
_, (_, past_key_values) = outputs.to_tuple()
_, past_key_values = outputs.to_tuple()
# create hypothetical multiple next token and extent to next_input_ids
next_tokens = tf.math.maximum(ids_tensor((self.batch_size, 3), config.vocab_size), 2)

7
tests/t5/test_modeling_tf_t5.py
View File

@@ -98,13 +98,10 @@ class TFT5ModelTester:
encoder_output = result.encoder_last_hidden_state
self.parent.assertListEqual(list(encoder_output.shape), [self.batch_size, self.seq_length, self.hidden_size])
self.parent.assertListEqual(list(decoder_output.shape), [self.batch_size, self.seq_length, self.hidden_size])
self.parent.assertEqual(len(decoder_past), 2)
# decoder_past[0] should correspond to encoder output
self.parent.assertTrue(tf.reduce_all(tf.math.equal(decoder_past[0][0], encoder_output)))
# There should be `num_layers` key value embeddings stored in decoder_past[1]
self.parent.assertEqual(len(decoder_past[1]), config.num_layers)
self.parent.assertEqual(len(decoder_past), config.num_layers)
# There should be a self attn key, a self attn value, a cross attn key and a cross attn value stored in each decoder_past[1] tuple
self.parent.assertEqual(len(decoder_past[1][0]), 4)
self.parent.assertEqual(len(decoder_past[0]), 4)
def create_and_check_t5_with_lm_head(self, config, input_ids, input_mask, token_labels):
model = TFT5ForConditionalGeneration(config=config)