tests/models/led/test_modeling_led.py - main - githubmirror/transformers

tests/models/led/test_modeling_led.py
584 строки94 KB
Ilyas Moutawwakil
[PoC] HF exporters (#41992)
02 июл 2026, 21:11
02 июл 2026, 21:11fde2b96
# Copyright 2021 Iz Beltagy, Matthew E. Peters, Arman Cohan and The HuggingFace Inc. 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.
"""Testing suite for the PyTorch LED model."""

import copy
import tempfile
import unittest
from functools import cached_property

from transformers import LEDConfig, is_torch_available
from transformers.models.auto import get_values
from transformers.testing_utils import (
    require_sentencepiece,
    require_tokenizers,
    require_torch,
    require_torch_fp16,
    slow,
    torch_device,
)

from ...generation.test_utils import GenerationTesterMixin
from ...test_configuration_common import ConfigTester
from ...test_modeling_common import ModelTesterMixin, ids_tensor
from ...test_pipeline_mixin import PipelineTesterMixin


if is_torch_available():
    import torch

    from transformers import (
        MODEL_FOR_QUESTION_ANSWERING_MAPPING,
        LEDForConditionalGeneration,
        LEDForQuestionAnswering,
        LEDModel,
        LEDTokenizer,
    )
    from transformers.models.led.modeling_led import LEDDecoder, LEDEncoder


def prepare_led_inputs_dict(
    config,
    input_ids,
    decoder_input_ids,
    attention_mask=None,
    decoder_attention_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)

    return {
        "input_ids": input_ids,
        "decoder_input_ids": decoder_input_ids,
        "attention_mask": attention_mask,
        "decoder_attention_mask": decoder_attention_mask,
    }


class LEDModelTester:
    def __init__(
        self,
        parent,
        batch_size=13,
        seq_length=11,
        is_training=True,
        use_labels=False,
        vocab_size=99,
        hidden_size=16,
        num_hidden_layers=2,
        num_attention_heads=4,
        intermediate_size=4,
        hidden_act="gelu",
        hidden_dropout_prob=0.1,
        attention_probs_dropout_prob=0.1,
        max_position_embeddings=32,
        eos_token_id=2,
        pad_token_id=1,
        bos_token_id=0,
        attention_window=4,
    ):
        self.parent = parent
        self.batch_size = batch_size
        self.seq_length = seq_length
        self.is_training = is_training
        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.eos_token_id = eos_token_id
        self.pad_token_id = pad_token_id
        self.bos_token_id = bos_token_id
        self.attention_window = attention_window

        # `ModelTesterMixin.test_attention_outputs` is expecting attention tensors to be of size
        # [num_attention_heads, encoder_seq_length, encoder_key_length], but LongformerSelfAttention
        # returns attention of shape [num_attention_heads, encoder_seq_length, self.attention_window + 1]
        # because its local attention only attends to `self.attention_window + 1` locations
        # (assuming no token with global attention, otherwise the last dimension of attentions
        # is x + self.attention_window + 1, where x is the number of tokens with global attention)
        # x is set to 1
        self.encoder_key_length = self.attention_window + 2

        # because of padding `encoder_seq_length`, is different from `seq_length`. Relevant for
        # the `test_attention_outputs` and `test_hidden_states_output` tests
        self.encoder_seq_length = self.seq_length

    def prepare_config_and_inputs(self):
        input_ids = ids_tensor([self.batch_size, self.seq_length], self.vocab_size)
        input_ids = ids_tensor([self.batch_size, self.seq_length], self.vocab_size).clamp(
            3,
        )
        input_ids[:, -1] = self.eos_token_id  # Eos Token

        decoder_input_ids = ids_tensor([self.batch_size, self.seq_length], self.vocab_size)

        config = self.get_config()
        inputs_dict = prepare_led_inputs_dict(config, input_ids, decoder_input_ids)
        return config, inputs_dict

    def get_config(self):
        return LEDConfig(
            vocab_size=self.vocab_size,
            d_model=self.hidden_size,
            encoder_layers=self.num_hidden_layers,
            decoder_layers=self.num_hidden_layers,
            encoder_attention_heads=self.num_attention_heads,
            decoder_attention_heads=self.num_attention_heads,
            encoder_ffn_dim=self.intermediate_size,
            decoder_ffn_dim=self.intermediate_size,
            dropout=self.hidden_dropout_prob,
            attention_dropout=self.attention_probs_dropout_prob,
            max_position_embeddings=self.max_position_embeddings,
            eos_token_id=self.eos_token_id,
            bos_token_id=self.bos_token_id,
            pad_token_id=self.pad_token_id,
            attention_window=self.attention_window,
        )

    def get_pipeline_config(self):
        config = self.get_config()
        config.max_position_embeddings = 100
        config.vocab_size = 300
        return config

    def prepare_config_and_inputs_for_common(self):
        config, inputs_dict = self.prepare_config_and_inputs()
        global_attention_mask = torch.zeros_like(inputs_dict["input_ids"])
        global_attention_mask[:, -1] = 1
        inputs_dict["global_attention_mask"] = global_attention_mask

        return config, inputs_dict

    def create_and_check_decoder_model_past_large_inputs(self, config, inputs_dict):
        model = LEDModel(config=config).get_decoder().to(torch_device).eval()
        input_ids = inputs_dict["input_ids"]
        attention_mask = inputs_dict["attention_mask"]

        # first forward pass
        outputs = model(input_ids, attention_mask=attention_mask, use_cache=True)

        output, past_key_values = outputs.to_tuple()

        # create hypothetical multiple next token and extent to next_input_ids
        next_tokens = ids_tensor((self.batch_size, 3), config.vocab_size)
        next_attn_mask = ids_tensor((self.batch_size, 3), 2)

        # append to next input_ids and
        next_input_ids = torch.cat([input_ids, next_tokens], dim=-1)
        next_attention_mask = torch.cat([attention_mask, next_attn_mask], dim=-1)

        output_from_no_past = model(next_input_ids, attention_mask=next_attention_mask)["last_hidden_state"]
        output_from_past = model(next_tokens, attention_mask=next_attention_mask, 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[:, -3:, random_slice_idx].detach()
        output_from_past_slice = output_from_past[:, :, random_slice_idx].detach()

        self.parent.assertTrue(output_from_past_slice.shape[1] == next_tokens.shape[1])

        # test that outputs are equal for slice
        self.parent.assertTrue(torch.allclose(output_from_past_slice, output_from_no_past_slice, atol=1e-2))

    def check_encoder_decoder_model_standalone(self, config, inputs_dict):
        model = LEDModel(config=config).to(torch_device).eval()
        outputs = model(**inputs_dict)

        encoder_last_hidden_state = outputs.encoder_last_hidden_state
        last_hidden_state = outputs.last_hidden_state

        with tempfile.TemporaryDirectory() as tmpdirname:
            encoder = model.get_encoder()
            encoder.save_pretrained(tmpdirname)
            encoder = LEDEncoder.from_pretrained(tmpdirname).to(torch_device)

        encoder_last_hidden_state_2 = encoder(
            inputs_dict["input_ids"],
            attention_mask=inputs_dict["attention_mask"],
            global_attention_mask=inputs_dict["global_attention_mask"],
        )[0]

        self.parent.assertTrue((encoder_last_hidden_state_2 - encoder_last_hidden_state).abs().max().item() < 1e-3)

        with tempfile.TemporaryDirectory() as tmpdirname:
            decoder = model.get_decoder()
            decoder.save_pretrained(tmpdirname)
            decoder = LEDDecoder.from_pretrained(tmpdirname).to(torch_device)

        last_hidden_state_2 = decoder(
            input_ids=inputs_dict["decoder_input_ids"],
            attention_mask=inputs_dict["decoder_attention_mask"],
            encoder_hidden_states=encoder_last_hidden_state,
            encoder_attention_mask=inputs_dict["attention_mask"],
        )[0]

        self.parent.assertTrue((last_hidden_state_2 - last_hidden_state).abs().max().item() < 1e-3)

    def check_global_attention(self, config, inputs_dict):
        model = LEDModel(config=config).to(torch_device).eval()
        model.config.output_attentions = True
        attention_mask = ids_tensor(inputs_dict["input_ids"].shape, vocab_size=2)
        global_attention_mask = torch.zeros_like(attention_mask)

        # set some tokens to global_attention
        num_tokens_with_global_attention = 2

        attention_mask[:, 2 : 2 + num_tokens_with_global_attention] = 1
        global_attention_mask[:, 2 : 2 + num_tokens_with_global_attention] = 1
        inputs_dict["attention_mask"] = attention_mask
        inputs_dict["global_attention_mask"] = global_attention_mask

        outputs = model(**inputs_dict)
        self.parent.assertIsNotNone(outputs.encoder_global_attentions)

        # setting `num_tokens_with_global_attention` to global_attentions yields
        # makes last dim to be of `num_tokens_with_global_attention`
        self.parent.assertTrue(
            outputs.encoder_global_attentions[0].shape,
            (self.batch_size, self.num_attention_heads, self.encoder_seq_length, num_tokens_with_global_attention),
        )


@require_torch
class LEDModelTest(ModelTesterMixin, GenerationTesterMixin, PipelineTesterMixin, unittest.TestCase):
    all_model_classes = (
        (LEDModel, LEDForConditionalGeneration, LEDForQuestionAnswering) if is_torch_available() else ()
    )
    pipeline_model_mapping = (
        {
            "feature-extraction": LEDModel,
        }
        if is_torch_available()
        else {}
    )
    is_encoder_decoder = True

    test_missing_keys = False
    test_torch_exportable = False  # chunked sliding-window attention is fundamentally data-dependent

    # TODO: Fix the failed tests when this model gets more usage
    def is_pipeline_test_to_skip(
        self,
        pipeline_test_case_name,
        config_class,
        model_architecture,
        tokenizer_name,
        image_processor_name,
        feature_extractor_name,
        processor_name,
    ):
        if pipeline_test_case_name == "QAPipelineTests" and not tokenizer_name.endswith("Fast"):
            return True

        return False

    def setUp(self):
        self.model_tester = LEDModelTester(self)
        self.config_tester = ConfigTester(self, config_class=LEDConfig)

    def test_config(self):
        self.config_tester.run_common_tests()

    def test_save_load_strict(self):
        config, inputs_dict = self.model_tester.prepare_config_and_inputs()
        for model_class in self.all_model_classes:
            model = model_class(config)

            with tempfile.TemporaryDirectory() as tmpdirname:
                model.save_pretrained(tmpdirname)
                model2, info = model_class.from_pretrained(tmpdirname, output_loading_info=True)
            self.assertEqual(info["missing_keys"], set())

    def test_decoder_model_past_with_large_inputs(self):
        config_and_inputs = self.model_tester.prepare_config_and_inputs()
        self.model_tester.create_and_check_decoder_model_past_large_inputs(*config_and_inputs)

    def test_encoder_decoder_model_standalone(self):
        config_and_inputs = self.model_tester.prepare_config_and_inputs_for_common()
        self.model_tester.check_encoder_decoder_model_standalone(*config_and_inputs)

    def test_global_attention(self):
        config_and_inputs = self.model_tester.prepare_config_and_inputs_for_common()
        self.model_tester.check_global_attention(*config_and_inputs)

    def prepare_config_and_inputs_for_generate(self, *args, **kwargs):
        config, inputs_dict = super().prepare_config_and_inputs_for_generate(*args, **kwargs)
        # LED computes attention scores based on mask indices if `is_global`
        inputs_dict.pop("global_attention_mask")
        return config, inputs_dict

    def test_inputs_embeds(self):
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()

        for model_class in self.all_model_classes:
            model = model_class(config)
            model.to(torch_device)
            model.eval()

            inputs = copy.deepcopy(self._prepare_for_class(inputs_dict, model_class))

            if not self.is_encoder_decoder:
                input_ids = inputs["input_ids"]
                del inputs["input_ids"]
            else:
                encoder_input_ids = inputs["input_ids"]
                decoder_input_ids = inputs.get("decoder_input_ids", encoder_input_ids)
                del inputs["input_ids"]
                inputs.pop("decoder_input_ids", None)

            wte = model.get_input_embeddings()
            if not self.is_encoder_decoder:
                inputs["inputs_embeds"] = wte(input_ids)
            else:
                inputs["inputs_embeds"] = wte(encoder_input_ids)
                inputs["decoder_inputs_embeds"] = wte(decoder_input_ids)

            with torch.no_grad():
                model(**inputs)[0]

    @require_torch_fp16
    def test_generate_fp16(self):
        config, input_dict = self.model_tester.prepare_config_and_inputs()
        input_ids = input_dict["input_ids"]
        attention_mask = input_ids.ne(1).to(torch_device)
        model = LEDForConditionalGeneration(config).eval().to(torch_device)
        model.half()
        model.generate(input_ids, attention_mask=attention_mask)
        model.generate(num_beams=4, do_sample=True, early_stopping=False, num_return_sequences=3)

    @unittest.skip(reason="Longformer cannot keep gradients in attentions or hidden states")
    def test_retain_grad_hidden_states_attentions(self):
        return

    def test_attention_outputs(self):
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
        config.return_dict = True

        seq_length = self.model_tester.seq_length
        encoder_seq_length = self.model_tester.encoder_seq_length
        encoder_key_length = self.model_tester.encoder_key_length

        for model_class in self.all_model_classes:
            inputs_dict["output_attentions"] = True
            inputs_dict["output_hidden_states"] = False
            config.return_dict = True
            model = model_class._from_config(config, attn_implementation="eager")
            config = model.config
            model.to(torch_device)
            model.eval()
            with torch.no_grad():
                outputs = model(**self._prepare_for_class(inputs_dict, model_class))
            attentions = outputs.encoder_attentions
            self.assertEqual(len(attentions), self.model_tester.num_hidden_layers)

            # check that output_attentions also work using config
            del inputs_dict["output_attentions"]
            config.output_attentions = True
            model = model_class(config)
            model.to(torch_device)
            model.eval()
            with torch.no_grad():
                outputs = model(**self._prepare_for_class(inputs_dict, model_class))
            attentions = outputs.encoder_attentions
            self.assertEqual(len(attentions), self.model_tester.num_hidden_layers)

            self.assertListEqual(
                list(attentions[0].shape[-3:]),
                [self.model_tester.num_attention_heads, encoder_seq_length, encoder_key_length],
            )
            out_len = len(outputs)

            # global attention outputs are added as well => so +1 here
            correct_outlen = 6

            # loss is at first position
            if "labels" in inputs_dict:
                correct_outlen += 1  # loss is added to beginning
            # Question Answering model returns start_logits and end_logits
            if model_class in get_values(MODEL_FOR_QUESTION_ANSWERING_MAPPING):
                correct_outlen += 1  # start_logits and end_logits instead of only 1 output
            if "past_key_values" in outputs:
                correct_outlen += 1  # past_key_values have been returned

            self.assertEqual(out_len, correct_outlen)

            # decoder attentions
            decoder_attentions = outputs.decoder_attentions
            self.assertIsInstance(decoder_attentions, (list, tuple))
            self.assertEqual(len(decoder_attentions), self.model_tester.num_hidden_layers)
            self.assertListEqual(
                list(decoder_attentions[0].shape[-3:]),
                [self.model_tester.num_attention_heads, seq_length, seq_length],
            )

            # cross attentions
            cross_attentions = outputs.cross_attentions
            self.assertIsInstance(cross_attentions, (list, tuple))
            self.assertEqual(len(cross_attentions), self.model_tester.num_hidden_layers)
            self.assertListEqual(
                list(cross_attentions[0].shape[-3:]),
                [
                    self.model_tester.num_attention_heads,
                    seq_length,
                    seq_length,
                ],
            )

    def _check_encoder_attention_for_generate(self, attentions, batch_size, config, prompt_length):
        # overwrite because LED does not have (bs, num_heads, seq_len, seq_len) shape
        encoder_expected_shape = (
            batch_size,
            config.num_attention_heads,
            prompt_length,
            self.model_tester.attention_window // 2 * 2 + 1,
        )
        self.assertIsInstance(attentions, tuple)
        self.assertListEqual(
            [layer_attentions.shape for layer_attentions in attentions],
            [encoder_expected_shape] * len(attentions),
        )


def assert_tensors_close(a, b, atol=1e-12, prefix=""):
    """If tensors have different shapes, different values or a and b are not both tensors, raise a nice Assertion error."""
    if a is None and b is None:
        return True
    try:
        if torch.allclose(a, b, atol=atol):
            return True
        raise Exception
    except Exception:
        pct_different = (torch.gt((a - b).abs(), atol)).float().mean().item()
        if a.numel() > 100:
            msg = f"tensor values are {pct_different:.1%} percent different."
        else:
            msg = f"{a} != {b}"
        if prefix:
            msg = prefix + ": " + msg
        raise AssertionError(msg)


def _long_tensor(tok_lst):
    return torch.tensor(tok_lst, dtype=torch.long, device=torch_device)


TOLERANCE = 1e-4


@require_torch
@require_sentencepiece
@require_tokenizers
@slow
class LEDModelIntegrationTests(unittest.TestCase):
    """All the below results were obtained with the original checkpoints and code
    base from https://github.com/allenai/longformer.
    IMPORTANT: Note that the original checkpoints include a `position_embeddings` "hack"
    and have to be cut to have the correct shape.
    See: https://github.com/huggingface/transformers/pull/9278#issue-544709661.
    """

    @cached_property
    def default_tokenizer(self):
        return LEDTokenizer.from_pretrained("allenai/led-base-16384")

    def test_inference_no_head(self):
        model = LEDModel.from_pretrained("allenai/led-base-16384").to(torch_device)

        # change to intended input
        input_ids = _long_tensor([512 * [0, 31414, 232, 328, 740, 1140, 12695, 69]])
        decoder_input_ids = _long_tensor([128 * [0, 31414, 232, 328, 740, 1140, 12695, 69]])
        inputs_dict = prepare_led_inputs_dict(model.config, input_ids, decoder_input_ids)
        with torch.no_grad():
            output = model(**inputs_dict).last_hidden_state
        expected_shape = torch.Size((1, 1024, 768))
        self.assertEqual(output.shape, expected_shape)
        # change to expected output here
        expected_slice = torch.tensor(
            [[2.3050, 2.8279, 0.6531], [-1.8457, -0.1455, -3.5661], [-1.0186, 0.4586, -2.2043]], device=torch_device
        )
        torch.testing.assert_close(output[0, :3, :3], expected_slice, rtol=TOLERANCE, atol=TOLERANCE)

    def test_inference_head(self):
        model = LEDForConditionalGeneration.from_pretrained("allenai/led-base-16384").to(torch_device)

        # change to intended input
        input_ids = _long_tensor([512 * [0, 31414, 232, 328, 740, 1140, 12695, 69]])
        decoder_input_ids = _long_tensor([128 * [0, 31414, 232, 328, 740, 1140, 12695, 69]])
        inputs_dict = prepare_led_inputs_dict(model.config, input_ids, decoder_input_ids)
        with torch.no_grad():
            output = model(**inputs_dict, use_cache=False).logits
        expected_shape = torch.Size((1, 1024, model.config.vocab_size))
        self.assertEqual(output.shape, expected_shape)
        # change to expected output here
        expected_slice = torch.tensor(
            [[33.6507, 6.4572, 16.8089], [5.8739, -2.4238, 11.2902], [-3.2139, -4.3149, 4.2783]], device=torch_device
        )
        torch.testing.assert_close(output[0, :3, :3], expected_slice, rtol=TOLERANCE, atol=TOLERANCE)

    def test_seq_to_seq_generation(self):
        # this test requires 16GB of RAM
        hf = LEDForConditionalGeneration.from_pretrained("allenai/led-large-16384-arxiv").to(torch_device)
        tok = LEDTokenizer.from_pretrained("allenai/led-large-16384-arxiv")

        ARTICLE_LEP = r"""the lep experiments at the resonance of @xmath1-boson have tested the standard model ( sm ) at quantum level , measuring the @xmath1-decay into fermion pairs with an accuracy … [Строка слишком длинная. Вы можете скачать файл]

        ARTICLE_MAGNET = r"""it is well known that the classical magnetoresistance ( mr ) in metals or semiconductors with a closed free electron fermi surface increases quadratically with increasing  … [Строка слишком длинная. Вы можете скачать файл]

        dct = tok(
            [ARTICLE_LEP, ARTICLE_MAGNET],
            max_length=6144,
            padding="max_length",
            truncation=True,
            return_tensors="pt",
        )

        hypotheses_batch = hf.generate(
            input_ids=dct["input_ids"].to(torch_device),
            attention_mask=dct["attention_mask"].to(torch_device),
            num_beams=4,
            max_length=512,
            early_stopping=True,
            no_repeat_ngram_size=3,
        )

        EXPECTED_LEP = (
            " the physics of @xmath0-boson will again play the central role in the frontier of particle physics if the"
            " gigaz option of the international linear collider ( ilc ) can be realized in its first phase . \n the"
            " expected sensitivity to the branching ratio of rare decays , especially its exotic or rare processes ,"
            " should be investigated comprehensively to evaluate their potential in probing new physics . in this work"
            " \n , we study the rare decay into light higgs boson(s ) in the framework of the minimal supersymmetric"
            " standard model ( mssm ) , where a light cp - odd higgs - boson with singlet - dominant component may"
            " naturally arise from the spontaneous breaking of some approximate global symmetry . "
        )

        EXPECTED_MAGNET = (
            " the recent experiment in the surface states of the topological insulator bi@xmath0se @xmath1 , however ,"
            " reported that a large positive magnetoresistance becomes very linear in perpendicular magnetic field"
            " even in an opposite situation where the carrier sheet density is high that all electrons occupy more"
            " than one landau levels . \n it is striking that this observation is in conflict with abrikosov s model"
            " and also with the classical parish - littlewood model . "
        )

        generated = tok.decode(hypotheses_batch.tolist(), skip_special_tokens=True)
        assert generated == [EXPECTED_LEP, EXPECTED_MAGNET]