# Copyright 2023 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 tempfile import unittest from transformers import T5Config, is_torch_available from transformers.testing_utils import ( require_sentencepiece, require_tokenizers, require_torch, slow, torch_device, ) from ...generation.test_utils import GenerationTesterMixin from ...test_modeling_common import ModelTesterMixin, ids_tensor from ...test_pipeline_mixin import PipelineTesterMixin if is_torch_available(): import torch from transformers import AutoTokenizer, UMT5ForConditionalGeneration, UMT5ForQuestionAnswering, UMT5Model # Copied from test.models.t5.test_modeling_t5.T5ModelTester with T5->UMT5 class UMT5ModelTester: def __init__( self, parent, vocab_size=99, batch_size=13, encoder_seq_length=7, decoder_seq_length=9, # For common tests is_training=True, use_attention_mask=True, use_labels=False, hidden_size=32, num_hidden_layers=5, num_attention_heads=4, d_ff=37, relative_attention_num_buckets=8, dropout_rate=0.1, initializer_factor=0.002, eos_token_id=1, pad_token_id=0, decoder_start_token_id=0, scope=None, decoder_layers=None, ): self.parent = parent self.batch_size = batch_size self.encoder_seq_length = encoder_seq_length self.decoder_seq_length = decoder_seq_length # For common tests self.seq_length = self.decoder_seq_length self.is_training = is_training self.use_attention_mask = use_attention_mask 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.d_ff = d_ff self.relative_attention_num_buckets = relative_attention_num_buckets self.dropout_rate = dropout_rate self.initializer_factor = initializer_factor self.eos_token_id = eos_token_id self.pad_token_id = pad_token_id self.decoder_start_token_id = decoder_start_token_id self.scope = None self.decoder_layers = decoder_layers def get_large_model_config(self): return T5Config.from_pretrained("google/umt5-base") def prepare_inputs_dict( self, config, input_ids, decoder_input_ids, attention_mask=None, decoder_attention_mask=None, head_mask=None, decoder_head_mask=None, cross_attn_head_mask=None, ): if attention_mask is None: attention_mask = input_ids.ne(config.pad_token_id) if decoder_attention_mask is None: decoder_attention_mask = decoder_input_ids.ne(config.pad_token_id) if head_mask is None: head_mask = torch.ones(config.num_hidden_layers, config.num_attention_heads, device=torch_device) if decoder_head_mask is None: decoder_head_mask = torch.ones(config.num_decoder_layers, config.num_attention_heads, device=torch_device) if cross_attn_head_mask is None: cross_attn_head_mask = torch.ones( config.num_decoder_layers, config.num_attention_heads, device=torch_device ) return { "input_ids": input_ids, "decoder_input_ids": decoder_input_ids, "attention_mask": attention_mask, "decoder_attention_mask": decoder_attention_mask, "head_mask": head_mask, "decoder_head_mask": decoder_head_mask, "cross_attn_head_mask": cross_attn_head_mask, } def prepare_config_and_inputs(self): input_ids = ids_tensor([self.batch_size, self.encoder_seq_length], self.vocab_size) decoder_input_ids = ids_tensor([self.batch_size, self.decoder_seq_length], self.vocab_size) # we need to clamp the input ids here to avoid having pad token in between # this is because for NllbMoe the position_ids are prepared such that # all pad tokens have pos id = 2 and rest are between 2..seq_length # and the seq_length here is seq_length - num_pad_tokens # but when using past, there is no way of knowing if the past input ids had # pad tokens in them, which results in incorrect seq_lenth and which in turn results in # position_ids being off by num_pad_tokens in past input input_ids = input_ids.clamp(self.pad_token_id + 1) decoder_input_ids = decoder_input_ids.clamp(self.pad_token_id + 1) config = self.get_config() config.encoder_attention_heads = config.num_attention_heads input_dict = self.prepare_inputs_dict(config, input_ids, decoder_input_ids) return config, input_dict def prepare_config_and_inputs_for_common(self): config, inputs_dict = self.prepare_config_and_inputs() return config, inputs_dict def get_pipeline_config(self): return T5Config( vocab_size=166, # t5 forces 100 extra tokens d_model=self.hidden_size, d_ff=self.d_ff, d_kv=self.hidden_size // self.num_attention_heads, num_layers=self.num_hidden_layers, num_decoder_layers=self.decoder_layers, num_heads=self.num_attention_heads, relative_attention_num_buckets=self.relative_attention_num_buckets, dropout_rate=self.dropout_rate, initializer_factor=self.initializer_factor, eos_token_id=self.eos_token_id, bos_token_id=self.pad_token_id, pad_token_id=self.pad_token_id, decoder_start_token_id=self.decoder_start_token_id, ) def get_config(self): return T5Config( vocab_size=self.vocab_size, d_model=self.hidden_size, d_ff=self.d_ff, d_kv=self.hidden_size // self.num_attention_heads, num_layers=self.num_hidden_layers, num_decoder_layers=self.decoder_layers, num_heads=self.num_attention_heads, relative_attention_num_buckets=self.relative_attention_num_buckets, dropout_rate=self.dropout_rate, initializer_factor=self.initializer_factor, eos_token_id=self.eos_token_id, bos_token_id=self.pad_token_id, pad_token_id=self.pad_token_id, decoder_start_token_id=self.decoder_start_token_id, ) def create_and_check_model( self, config, input_ids, decoder_input_ids, attention_mask, decoder_attention_mask, lm_labels, ): model = UMT5Model(config=config) model.to(torch_device) model.eval() result = model( input_ids=input_ids, decoder_input_ids=decoder_input_ids, attention_mask=attention_mask, decoder_attention_mask=decoder_attention_mask, ) result = model(input_ids=input_ids, decoder_input_ids=decoder_input_ids) decoder_output = result.last_hidden_state decoder_past = result.past_key_values encoder_output = result.encoder_last_hidden_state self.parent.assertEqual(encoder_output.size(), (self.batch_size, self.encoder_seq_length, self.hidden_size)) self.parent.assertEqual(decoder_output.size(), (self.batch_size, self.decoder_seq_length, self.hidden_size)) # There should be `num_layers` key value embeddings stored in decoder_past 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 tuple self.parent.assertEqual(len(decoder_past[0]), 4) def create_and_check_decoder_model_past( self, config, input_ids, decoder_input_ids, attention_mask, decoder_attention_mask, lm_labels, ): model = UMT5Model(config=config).get_decoder().to(torch_device).eval() # first forward pass outputs = model(input_ids, use_cache=True) outputs_use_cache_conf = model(input_ids) outputs_no_past = model(input_ids, use_cache=False) self.parent.assertTrue(len(outputs) == len(outputs_use_cache_conf)) self.parent.assertTrue(len(outputs) == len(outputs_no_past) + 1) output, past_key_values = outputs.to_tuple() # create hypothetical next token and extent to next_input_ids next_tokens = ids_tensor((self.batch_size, 1), config.vocab_size) # append to next input_ids and next_input_ids = torch.cat([input_ids, next_tokens], dim=-1) output_from_no_past = model(next_input_ids)["last_hidden_state"] output_from_past = model(next_tokens, past_key_values=past_key_values)["last_hidden_state"] # select random slice random_slice_idx = ids_tensor((1,), output_from_past.shape[-1]).item() output_from_no_past_slice = output_from_no_past[:, -1, random_slice_idx].detach() output_from_past_slice = output_from_past[:, 0, random_slice_idx].detach() # test that outputs are equal for slice self.parent.assertTrue(torch.allclose(output_from_past_slice, output_from_no_past_slice, atol=1e-3)) def create_and_check_model_fp16_forward( self, config, input_dict, ): model = UMT5Model(config=config).to(torch_device).half().eval() output = model(**input_dict)["last_hidden_state"] self.parent.assertFalse(torch.isnan(output).any().item()) @require_torch class UMT5ModelTest(ModelTesterMixin, GenerationTesterMixin, PipelineTesterMixin, unittest.TestCase): all_model_classes = ( (UMT5Model, UMT5ForConditionalGeneration, UMT5ForQuestionAnswering) if is_torch_available() else () ) all_generative_model_classes = (UMT5ForConditionalGeneration,) if is_torch_available() else () pipeline_model_mapping = ( { "conversational": UMT5ForConditionalGeneration, "feature-extraction": UMT5Model, "summarization": UMT5ForConditionalGeneration, "text2text-generation": UMT5ForConditionalGeneration, "translation": UMT5ForConditionalGeneration, "question-answering": UMT5ForQuestionAnswering, } if is_torch_available() else {} ) is_encoder_decoder = True fx_compatible = False test_pruning = False test_missing_keys = True test_torchscript = True # The small UMT5 model needs higher percentages for CPU/MP tests model_split_percents = [0.8, 0.9] def setUp(self): self.model_tester = UMT5ModelTester(self) @unittest.skip("Test has a segmentation fault on torch 1.8.0") def test_export_to_onnx(self): config_and_inputs = self.model_tester.prepare_config_and_inputs() model = UMT5Model(config_and_inputs[0]).to(torch_device) with tempfile.TemporaryDirectory() as tmpdirname: torch.onnx.export( model, (config_and_inputs[1], config_and_inputs[3], config_and_inputs[2]), f"{tmpdirname}/t5_test.onnx", export_params=True, opset_version=9, input_names=["input_ids", "decoder_input_ids"], ) @unittest.skipIf(torch_device == "cpu", "Cant do half precision") def test_model_fp16_forward(self): config_and_inputs = self.model_tester.prepare_config_and_inputs() self.model_tester.create_and_check_model_fp16_forward(*config_and_inputs) def test_generate_with_head_masking(self): attention_names = ["encoder_attentions", "decoder_attentions", "cross_attentions"] config_and_inputs = self.model_tester.prepare_config_and_inputs() config = config_and_inputs[0] model = UMT5ForConditionalGeneration(config).eval() model.to(torch_device) head_masking = { "head_mask": torch.zeros(config.num_layers, config.num_heads, device=torch_device), "decoder_head_mask": torch.zeros(config.num_decoder_layers, config.num_heads, device=torch_device), "cross_attn_head_mask": torch.zeros(config.num_decoder_layers, config.num_heads, device=torch_device), } for attn_name, (name, mask) in zip(attention_names, head_masking.items()): head_masks = {name: mask} # Explicitly pass decoder_head_mask as it is required from T5 model when head_mask specified if name == "head_mask": head_masks["decoder_head_mask"] = torch.ones( config.num_decoder_layers, config.num_heads, device=torch_device ) out = model.generate( config_and_inputs[1]["input_ids"], num_beams=1, max_length=3, output_attentions=True, return_dict_in_generate=True, **head_masks, ) # We check the state of decoder_attentions and cross_attentions just from the last step attn_weights = out[attn_name] if attn_name == attention_names[0] else out[attn_name][-1] self.assertEqual(sum([w.sum().item() for w in attn_weights]), 0.0) @unittest.skip("Does not work on the tiny model as we keep hitting edge cases.") def test_disk_offload(self): pass @require_torch @require_sentencepiece @require_tokenizers class Umt5IntegrationTest(unittest.TestCase): @slow def test_small_integration_test(self): """ For comparison run the kaggle notbook available here : https://www.kaggle.com/arthurzucker/umt5-inference """ model = UMT5ForConditionalGeneration.from_pretrained("google/umt5-small", return_dict=True).to(torch_device) tokenizer = AutoTokenizer.from_pretrained("google/umt5-small", use_fast=False) input_text = [ "Bonjour monsieur bien .", "No se como puedo .", "This is the reason why we them.", "The walks in , seats", "A walks into a bar and orders a with pinch of .", ] input_ids = tokenizer(input_text, return_tensors="pt", padding=True).input_ids # fmt: off EXPECTED_IDS = torch.tensor( [ [ 38530, 210703, 256299, 1410, 256298, 274, 1, 0,0, 0, 0, 0, 0, 0, 0, 0,0, 0], [ 826, 321, 671, 25922, 256299, 274, 1, 0,0, 0, 0, 0, 0, 0, 0, 0,0, 0], [ 1460, 339, 312, 19014, 10620, 758, 256299, 2355,274, 1, 0, 0, 0, 0, 0, 0,0, 0], [ 517, 256299, 14869, 281, 301, 256298, 275, 119983,1, 0, 0, 0, 0, 0, 0, 0,0, 0], [ 320, 256299, 14869, 281, 2234, 289, 2275, 333,61391, 289, 256298, 543, 256297, 168714, 329, 256296,274, 1], ] ) # fmt: on self.assertEqual(input_ids, EXPECTED_IDS) generated_ids = model.generate(input_ids.to(torch_device)) EXPECTED_FILLING = [ " et [eod] .. [eod] πŸ’ πŸ’ πŸ’ πŸ’ πŸ’ πŸ’ πŸ’ πŸ’ πŸ’ πŸ’ πŸ’ ajΕ‘ietostolleuxajΕ‘ie", "..,<0x0A>...spech <0x0A> ", " are not going to be a part of the world. We are not going to be a part of and<0x0A>.", " door, the door ν”Όν•΄[/", "nyone who drink a alcohol A A. This I", ] filling = tokenizer.batch_decode(generated_ids) self.assertTrue(filling, EXPECTED_FILLING)