# Copyright 2021 The HuggingFace Team. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import unittest import numpy as np from transformers import GPT2Config, GPT2Tokenizer, is_flax_available from transformers.testing_utils import require_flax, slow from ...test_modeling_flax_common import FlaxModelTesterMixin, floats_tensor, ids_tensor, random_attention_mask if is_flax_available(): import jax import jax.numpy as jnp from transformers.models.gpt2.modeling_flax_gpt2 import FlaxGPT2LMHeadModel, FlaxGPT2Model class FlaxGPT2ModelTester: def __init__( self, parent, batch_size=14, seq_length=7, is_training=True, use_input_mask=True, use_token_type_ids=False, use_labels=True, vocab_size=99, hidden_size=32, num_hidden_layers=2, num_attention_heads=4, intermediate_size=37, hidden_act="gelu", hidden_dropout_prob=0.1, attention_probs_dropout_prob=0.1, max_position_embeddings=512, initializer_range=0.02, ): self.parent = parent self.batch_size = batch_size self.seq_length = seq_length self.is_training = is_training self.use_input_mask = use_input_mask self.use_token_type_ids = use_token_type_ids self.use_labels = use_labels self.vocab_size = vocab_size self.hidden_size = hidden_size self.num_hidden_layers = num_hidden_layers self.num_attention_heads = num_attention_heads self.intermediate_size = intermediate_size self.hidden_act = hidden_act self.hidden_dropout_prob = hidden_dropout_prob self.attention_probs_dropout_prob = attention_probs_dropout_prob self.max_position_embeddings = max_position_embeddings self.initializer_range = initializer_range self.scope = None self.bos_token_id = vocab_size - 1 self.eos_token_id = vocab_size - 1 self.pad_token_id = vocab_size - 1 def prepare_config_and_inputs(self): input_ids = ids_tensor([self.batch_size, self.seq_length], self.vocab_size) input_mask = None if self.use_input_mask: input_mask = random_attention_mask([self.batch_size, self.seq_length]) config = GPT2Config( vocab_size=self.vocab_size, n_embd=self.hidden_size, n_layer=self.num_hidden_layers, n_head=self.num_attention_heads, n_positions=self.max_position_embeddings, use_cache=False, bos_token_id=self.bos_token_id, eos_token_id=self.eos_token_id, pad_token_id=self.pad_token_id, ) return (config, input_ids, input_mask) def prepare_config_and_inputs_for_common(self): config_and_inputs = self.prepare_config_and_inputs() config, input_ids, attention_mask = config_and_inputs inputs_dict = {"input_ids": input_ids, "attention_mask": attention_mask} return config, inputs_dict def prepare_config_and_inputs_for_decoder(self): config, input_ids, attention_mask = self.prepare_config_and_inputs() encoder_hidden_states = floats_tensor([self.batch_size, self.seq_length, self.hidden_size]) encoder_attention_mask = ids_tensor([self.batch_size, self.seq_length], vocab_size=2) return ( config, input_ids, attention_mask, encoder_hidden_states, encoder_attention_mask, ) def check_use_cache_forward(self, model_class_name, config, input_ids, attention_mask): max_decoder_length = 20 model = model_class_name(config) past_key_values = model.init_cache(input_ids.shape[0], max_decoder_length) attention_mask = jnp.ones((input_ids.shape[0], max_decoder_length), dtype="i4") position_ids = jnp.broadcast_to( jnp.arange(input_ids.shape[-1] - 1)[None, :], (input_ids.shape[0], input_ids.shape[-1] - 1) ) outputs_cache = model( input_ids[:, :-1], attention_mask=attention_mask, past_key_values=past_key_values, position_ids=position_ids, ) position_ids = jnp.array(input_ids.shape[0] * [[input_ids.shape[-1] - 1]], dtype="i4") outputs_cache_next = model( input_ids[:, -1:], attention_mask=attention_mask, past_key_values=outputs_cache.past_key_values, position_ids=position_ids, ) outputs = model(input_ids) diff = np.max(np.abs((outputs_cache_next[0][:, -1, :5] - outputs[0][:, -1, :5]))) self.parent.assertTrue(diff < 1e-3, msg=f"Max diff is {diff}") def check_use_cache_forward_with_attn_mask(self, model_class_name, config, input_ids, attention_mask): max_decoder_length = 20 model = model_class_name(config) attention_mask_cache = jnp.concatenate( [attention_mask, jnp.zeros((attention_mask.shape[0], max_decoder_length - attention_mask.shape[1]))], axis=-1, ) past_key_values = model.init_cache(input_ids.shape[0], max_decoder_length) position_ids = jnp.broadcast_to( jnp.arange(input_ids.shape[-1] - 1)[None, :], (input_ids.shape[0], input_ids.shape[-1] - 1) ) outputs_cache = model( input_ids[:, :-1], attention_mask=attention_mask_cache, past_key_values=past_key_values, position_ids=position_ids, ) position_ids = jnp.array(input_ids.shape[0] * [[input_ids.shape[-1] - 1]], dtype="i4") outputs_cache_next = model( input_ids[:, -1:], past_key_values=outputs_cache.past_key_values, attention_mask=attention_mask_cache, position_ids=position_ids, ) outputs = model(input_ids, attention_mask=attention_mask) diff = np.max(np.abs((outputs_cache_next[0][:, -1, :5] - outputs[0][:, -1, :5]))) self.parent.assertTrue(diff < 1e-3, msg=f"Max diff is {diff}") def check_bool_attention_mask_in_generation(self, model_class_name, config, input_ids, attention_mask): model = model_class_name(config) output_int_att_mask = model.generate( input_ids=input_ids, attention_mask=attention_mask, max_new_tokens=3, ) output_bool_att_mask = model.generate( input_ids=input_ids, attention_mask=attention_mask.astype(bool), max_new_tokens=3, ) self.parent.assertTrue( (output_bool_att_mask.sequences == output_int_att_mask.sequences).all(), "Generated response differ between boolean and integer attention mask", ) @require_flax class FlaxGPT2ModelTest(FlaxModelTesterMixin, unittest.TestCase): all_model_classes = (FlaxGPT2Model, FlaxGPT2LMHeadModel) if is_flax_available() else () def setUp(self): self.model_tester = FlaxGPT2ModelTester(self) def test_use_cache_forward(self): for model_class_name in self.all_model_classes: config, input_ids, attention_mask = self.model_tester.prepare_config_and_inputs() self.model_tester.check_use_cache_forward(model_class_name, config, input_ids, attention_mask) def test_use_cache_forward_with_attn_mask(self): for model_class_name in self.all_model_classes: config, input_ids, attention_mask = self.model_tester.prepare_config_and_inputs() self.model_tester.check_use_cache_forward_with_attn_mask( model_class_name, config, input_ids, attention_mask ) def test_bool_attention_mask_in_generation(self): for model_class_name in self.all_generative_model_classes: config, input_ids, attention_mask = self.model_tester.prepare_config_and_inputs() self.model_tester.check_bool_attention_mask_in_generation( model_class_name, config, input_ids, attention_mask ) @slow def test_batch_generation(self): tokenizer = GPT2Tokenizer.from_pretrained("openai-community/gpt2", pad_token="", padding_side="left") inputs = tokenizer(["Hello this is a long string", "Hey"], return_tensors="np", padding=True, truncation=True) model = FlaxGPT2LMHeadModel.from_pretrained("openai-community/gpt2") model.do_sample = False model.config.pad_token_id = model.config.eos_token_id jit_generate = jax.jit(model.generate) output_sequences = jit_generate(inputs["input_ids"], attention_mask=inputs["attention_mask"]).sequences output_string = tokenizer.batch_decode(output_sequences, skip_special_tokens=True) expected_string = [ "Hello this is a long string of words. I'm going to start with the first one.\n", "Hey, I'm not sure if I'm going to be able to do", ] self.assertListEqual(output_string, expected_string) @slow def test_model_from_pretrained(self): for model_class_name in self.all_model_classes: model = model_class_name.from_pretrained("openai-community/gpt2", from_pt=True) outputs = model(np.ones((1, 1))) self.assertIsNotNone(outputs)