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

TF XLA greedy generation (#15786)

Browse Source 
* First attempt at TF XLA generation

* Fix comments

* Update XLA greedy generate with direct XLA calls

* Support attention mask, prepare_inputs_for_generation no longer hardcoded for greedy

* Handle position_ids correctly

* make xla generate work for non xla case

* force using xla generate

* refactor

* more fixes

* finish cleaning

* finish

* finish

* clean gpt2 tests

* add gpt2 tests

* correct more cases

* up

* finish

* finish

* more fixes

* flake 8 stuff

* final rag fix

* Update src/transformers/models/rag/modeling_tf_rag.py

* finish t5 as well

* finish

* Update src/transformers/generation_utils.py

Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>
This commit is contained in:
Matt
2022-03-15 13:19:20 +00:00
committed by GitHub
parent e5bc438cc8
commit cd4c5c9060
8 changed files with 370 additions and 93 deletions
Download Patch File Download Diff File Expand all files Collapse all files

42
tests/t5/test_modeling_tf_t5.py
View File

@@ -227,6 +227,23 @@ class TFT5ModelTester:
# test that outputs are equal for slice
tf.debugging.assert_near(output_from_past_slice, output_from_no_past_slice, rtol=1e-3)
def create_and_check_t5_xla_generate(self, config, input_ids, *args):
config.eos_token_id = None
config.max_length = 10
config.do_sample = False
config.num_beams = 1
model = TFT5ForConditionalGeneration(config=config)
# make sure there are no pad tokens in prompt
input_ids = tf.where(input_ids != config.pad_token_id, input_ids, config.pad_token_id + 5)
generated = model.generate(input_ids)
generate_xla = tf.function(model.generate, jit_compile=True)
generated_xla = generate_xla(input_ids)
self.parent.assertListEqual(generated.numpy().tolist(), generated_xla.numpy().tolist())
def prepare_config_and_inputs_for_common(self):
config_and_inputs = self.prepare_config_and_inputs()
(config, input_ids, input_mask, token_labels) = config_and_inputs
@@ -280,6 +297,10 @@ class TFT5ModelTest(TFModelTesterMixin, unittest.TestCase):
config_and_inputs = self.model_tester.prepare_config_and_inputs()
self.model_tester.create_and_check_t5_decoder_model_past_large_inputs(*config_and_inputs)
def test_t5_model_xla_generate(self):
config_and_inputs = self.model_tester.prepare_config_and_inputs()
self.model_tester.create_and_check_t5_xla_generate(*config_and_inputs)
def test_model_common_attributes(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
@@ -454,6 +475,27 @@ class TFT5EncoderOnlyModelTest(TFModelTesterMixin, unittest.TestCase):
@require_sentencepiece
@require_tokenizers
class TFT5GenerationIntegrationTests(unittest.TestCase):
@slow
def test_greedy_xla_generate_simple(self):
model = TFT5ForConditionalGeneration.from_pretrained("t5-small")
tokenizer = T5Tokenizer.from_pretrained("t5-small")
sentence = "Translate English to German: Today is a beautiful day."
input_ids = tokenizer(sentence, return_tensors="tf", padding=True).input_ids
xla_generate = tf.function(model.generate, jit_compile=True)
output_ids = model.generate(input_ids)
output_ids_xla = xla_generate(input_ids)
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 = ["Heute ist ein schöner Tag."]
self.assertListEqual(expected_output_string, output_strings)
self.assertListEqual(expected_output_string, output_strings_xla)
@slow
def test_greedy_generate(self):
model = TFT5ForConditionalGeneration.from_pretrained("t5-small")