LLM caching breaks with shared-prefix labels under certain conditions about skorch HOT 3 CLOSED

    @pytest.fixture(scope='class')
    def model_cacheable(self):
        from transformers import AutoModelForCausalLM
        return AutoModelForCausalLM.from_pretrained('gpt2')

    def test_caching_works_shared_label_prefix_without_eos(self, model_cacheable, classifier_cls):
        from transformers import AutoTokenizer
        tokenizer = AutoTokenizer.from_pretrained(model_cacheable.config._name_or_path)
        tokenizer.add_eos_token = False

        generate_fn = model_cacheable.generate

        def patched_generate(*args, **kwargs):
            original_result = generate_fn(*args, **kwargs)
            no_eos_result = original_result[:, :-1]
            return no_eos_result

        with patch.object(model_cacheable, 'generate', new=patched_generate):
            clf = classifier_cls(model=model_cacheable, tokenizer=tokenizer)

            X = np.array([["Hey there"], ["No thank you"], ["Whatever"]])
            y = ['0', '1', '12']

            clf.fit(X, y)

            for label_ids in clf.label_ids_:
                assert label_ids[-1] != model_cacheable.config.eos_token_id

            clf.predict(X)
            clf.predict(X)

    def test_caching_works_shared_label_prefix_without_eos(self, model_cacheable, classifier_cls):
        clf = classifier_cls('HuggingFaceH4/zephyr-7b-beta')

        X = np.array([["Hey there"], ["No thank you"], ["Whatever"]])
        y = ['0', '1', '12']

        clf.fit(X, y)

        for label_ids in clf.label_ids_:
            assert label_ids[-1] != model_cacheable.config.eos_token_id

        clf.predict(X)
        clf.predict(X)