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

TF: XLA beam search + most generation-compatible models are now also XLA-generate-compatible (#17857)

Browse Source 
* working beam search 🎉

* XLA generation compatible with ALL classes

* add xla generation slow test
This commit is contained in:
Joao Gante
2022-06-29 12:41:01 +01:00
committed by GitHub
parent b8142753f9
commit e6d27ca5c8
16 changed files with 356 additions and 301 deletions
Download Patch File Download Diff File Expand all files Collapse all files

43
tests/models/gpt2/test_modeling_tf_gpt2.py
View File

@@ -294,21 +294,6 @@ class TFGPT2ModelTester:
result = model(inputs)
self.parent.assertEqual(result.logits.shape, (self.batch_size, self.seq_length, self.vocab_size))
def create_and_check_gpt2_xla_generate_fast(self, config, input_ids, *args):
config.eos_token_id = None # Generate until max length
config.max_length = 10
model = TFGPT2LMHeadModel(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 - 1)
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 create_and_check_gpt2_double_head(
self, config, input_ids, input_mask, head_mask, token_type_ids, mc_token_ids, *args
):
@@ -408,10 +393,6 @@ class TFGPT2ModelTest(TFModelTesterMixin, TFCoreModelTesterMixin, unittest.TestC
config_and_inputs = self.model_tester.prepare_config_and_inputs()
self.model_tester.create_and_check_gpt2_lm_head(*config_and_inputs)
def test_gpt2_xla_generate_fast(self):
config_and_inputs = self.model_tester.prepare_config_and_inputs()
self.model_tester.create_and_check_gpt2_xla_generate_fast(*config_and_inputs)
def test_gpt2_double_head(self):
config_and_inputs = self.model_tester.prepare_config_and_inputs()
self.model_tester.create_and_check_gpt2_double_head(*config_and_inputs)
@@ -653,3 +634,27 @@ class TFGPT2ModelLanguageGenerationTest(unittest.TestCase):
output_ids = xla_generate(**input_ids, do_sample=True, seed=[7, 0])
output_strings = tokenizer.batch_decode(output_ids, skip_special_tokens=True)
self.assertListEqual(output_strings, expected_output_string_xla)
@slow
def test_lm_generate_gpt2_beam_search_xla(self):
model = TFGPT2LMHeadModel.from_pretrained("gpt2")
tokenizer = GPT2Tokenizer.from_pretrained("gpt2")
tokenizer.pad_token = tokenizer.eos_token
tokenizer.padding_side = "left"
sentences = ["The dog", "The flying machine"]
expected_output_strings = [
"The dog was found in the backyard of a home in the 6500 block of South Main Street",
"The flying machine is a very powerful machine, but it's not a very powerful machine. It's",
]
input_ids = tokenizer(sentences, return_tensors="tf", padding=True)
output_ids = model.generate(**input_ids, do_sample=False, num_beams=2)
output_strings = tokenizer.batch_decode(output_ids, skip_special_tokens=True)
self.assertListEqual(output_strings, expected_output_strings)
xla_generate = tf.function(model.generate, jit_compile=True)
output_ids = xla_generate(**input_ids, do_sample=False, num_beams=2)
output_strings = tokenizer.batch_decode(output_ids, skip_special_tokens=True)
self.assertListEqual(output_strings, expected_output_strings)