transformers/tests/models/stablelm/test_modeling_stablelm.py

# coding=utf-8
# Copyright 2024 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 StableLm model. """


import unittest

from parameterized import parameterized

from transformers import StableLmConfig, is_torch_available, set_seed
from transformers.testing_utils import (
    require_bitsandbytes,
    require_flash_attn,
    require_torch,
    require_torch_sdpa,
    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 (
        AutoTokenizer,
        StableLmForCausalLM,
        StableLmForSequenceClassification,
        StableLmModel,
    )


# Copied from transformers.tests.models.persimmon.test_modeling_persimmon.PersimmonModelTester with Persimmon -> StableLm
class StableLmModelTester:
    # Ignore copy
    def __init__(
        self,
        parent,
        batch_size=13,
        seq_length=7,
        is_training=True,
        use_input_mask=True,
        use_token_type_ids=False,
        use_labels=True,
        vocab_size=99,
        hidden_size=64,
        num_hidden_layers=2,
        num_attention_heads=4,
        num_key_value_heads=4,
        intermediate_size=37,
        hidden_act="gelu",
        hidden_dropout_prob=0.1,
        attention_probs_dropout_prob=0.1,
        max_position_embeddings=512,
        type_vocab_size=16,
        type_sequence_label_size=2,
        initializer_range=0.02,
        num_labels=3,
        num_choices=4,
        pad_token_id=0,
        scope=None,
    ):
        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.num_key_value_heads = num_key_value_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.type_vocab_size = type_vocab_size
        self.type_sequence_label_size = type_sequence_label_size
        self.initializer_range = initializer_range
        self.num_labels = num_labels
        self.num_choices = num_choices
        self.pad_token_id = pad_token_id
        self.scope = scope

    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 = torch.tril(torch.ones(self.batch_size, self.seq_length)).to(torch_device)

        token_type_ids = None
        if self.use_token_type_ids:
            token_type_ids = ids_tensor([self.batch_size, self.seq_length], self.type_vocab_size)

        sequence_labels = None
        token_labels = None
        choice_labels = None
        if self.use_labels:
            sequence_labels = ids_tensor([self.batch_size], self.type_sequence_label_size)
            token_labels = ids_tensor([self.batch_size, self.seq_length], self.num_labels)
            choice_labels = ids_tensor([self.batch_size], self.num_choices)

        config = self.get_config()

        return config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels

    def get_config(self):
        return StableLmConfig(
            vocab_size=self.vocab_size,
            hidden_size=self.hidden_size,
            num_hidden_layers=self.num_hidden_layers,
            num_attention_heads=self.num_attention_heads,
            num_key_value_heads=self.num_key_value_heads,
            intermediate_size=self.intermediate_size,
            hidden_act=self.hidden_act,
            hidden_dropout_prob=self.hidden_dropout_prob,
            attention_probs_dropout_prob=self.attention_probs_dropout_prob,
            max_position_embeddings=self.max_position_embeddings,
            type_vocab_size=self.type_vocab_size,
            is_decoder=False,
            initializer_range=self.initializer_range,
            pad_token_id=self.pad_token_id,
        )

    def create_and_check_model(
        self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
    ):
        model = StableLmModel(config=config)
        model.to(torch_device)
        model.eval()
        result = model(input_ids, attention_mask=input_mask)
        result = model(input_ids)
        self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))

    def create_and_check_model_as_decoder(
        self,
        config,
        input_ids,
        token_type_ids,
        input_mask,
        sequence_labels,
        token_labels,
        choice_labels,
        encoder_hidden_states,
        encoder_attention_mask,
    ):
        config.add_cross_attention = True
        model = StableLmModel(config)
        model.to(torch_device)
        model.eval()
        result = model(
            input_ids,
            attention_mask=input_mask,
            encoder_hidden_states=encoder_hidden_states,
            encoder_attention_mask=encoder_attention_mask,
        )
        result = model(
            input_ids,
            attention_mask=input_mask,
            encoder_hidden_states=encoder_hidden_states,
        )
        result = model(input_ids, attention_mask=input_mask)
        self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))

    def create_and_check_for_causal_lm(
        self,
        config,
        input_ids,
        token_type_ids,
        input_mask,
        sequence_labels,
        token_labels,
        choice_labels,
        encoder_hidden_states,
        encoder_attention_mask,
    ):
        model = StableLmForCausalLM(config=config)
        model.to(torch_device)
        model.eval()
        result = model(input_ids, attention_mask=input_mask, labels=token_labels)
        self.parent.assertEqual(result.logits.shape, (self.batch_size, self.seq_length, self.vocab_size))

    def create_and_check_decoder_model_past_large_inputs(
        self,
        config,
        input_ids,
        token_type_ids,
        input_mask,
        sequence_labels,
        token_labels,
        choice_labels,
        encoder_hidden_states,
        encoder_attention_mask,
    ):
        config.is_decoder = True
        config.add_cross_attention = True
        model = StableLmForCausalLM(config=config)
        model.to(torch_device)
        model.eval()

        # first forward pass
        outputs = model(
            input_ids,
            attention_mask=input_mask,
            encoder_hidden_states=encoder_hidden_states,
            encoder_attention_mask=encoder_attention_mask,
            use_cache=True,
        )
        past_key_values = outputs.past_key_values

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

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

        output_from_no_past = model(
            next_input_ids,
            attention_mask=next_attention_mask,
            encoder_hidden_states=encoder_hidden_states,
            encoder_attention_mask=encoder_attention_mask,
            output_hidden_states=True,
        )["hidden_states"][0]
        output_from_past = model(
            next_tokens,
            attention_mask=next_attention_mask,
            encoder_hidden_states=encoder_hidden_states,
            encoder_attention_mask=encoder_attention_mask,
            past_key_values=past_key_values,
            output_hidden_states=True,
        )["hidden_states"][0]

        # 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-3))

    def prepare_config_and_inputs_for_common(self):
        config_and_inputs = self.prepare_config_and_inputs()
        (
            config,
            input_ids,
            token_type_ids,
            input_mask,
            sequence_labels,
            token_labels,
            choice_labels,
        ) = config_and_inputs
        inputs_dict = {"input_ids": input_ids, "attention_mask": input_mask}
        return config, inputs_dict


@require_torch
# Copied from transformers.tests.persimmon.test_modeling_persimmon.PersimmonModelTest with Persimmon -> StableLm
class StableLmModelTest(ModelTesterMixin, GenerationTesterMixin, PipelineTesterMixin, unittest.TestCase):
    all_model_classes = (
        (StableLmModel, StableLmForCausalLM, StableLmForSequenceClassification) if is_torch_available() else ()
    )
    pipeline_model_mapping = (
        {
            "feature-extraction": StableLmModel,
            "text-classification": StableLmForSequenceClassification,
            # TODO (ydshieh): check why these two fail. Fix them or skip them in a better way.
            # "text-generation": StableLmForCausalLM,
            # "zero-shot": StableLmForSequenceClassification,
        }
        if is_torch_available()
        else {}
    )

    all_generative_model_classes = (StableLmForCausalLM,) if is_torch_available() else ()
    test_headmasking = False
    test_pruning = False

    def setUp(self):
        self.model_tester = StableLmModelTester(self)
        self.config_tester = ConfigTester(self, config_class=StableLmConfig, hidden_size=37)

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

    def test_model(self):
        config_and_inputs = self.model_tester.prepare_config_and_inputs()
        self.model_tester.create_and_check_model(*config_and_inputs)

    def test_stablelm_sequence_classification_model(self):
        config, input_dict = self.model_tester.prepare_config_and_inputs_for_common()
        config.num_labels = 3
        input_ids = input_dict["input_ids"]
        attention_mask = input_ids.ne(1).to(torch_device)
        sequence_labels = ids_tensor([self.model_tester.batch_size], self.model_tester.type_sequence_label_size)
        model = StableLmForSequenceClassification(config)
        model.to(torch_device)
        model.eval()
        result = model(input_ids, attention_mask=attention_mask, labels=sequence_labels)
        self.assertEqual(result.logits.shape, (self.model_tester.batch_size, self.model_tester.num_labels))

    def test_stablelm_sequence_classification_model_for_single_label(self):
        config, input_dict = self.model_tester.prepare_config_and_inputs_for_common()
        config.num_labels = 3
        config.problem_type = "single_label_classification"
        input_ids = input_dict["input_ids"]
        attention_mask = input_ids.ne(1).to(torch_device)
        sequence_labels = ids_tensor([self.model_tester.batch_size], self.model_tester.type_sequence_label_size)
        model = StableLmForSequenceClassification(config)
        model.to(torch_device)
        model.eval()
        result = model(input_ids, attention_mask=attention_mask, labels=sequence_labels)
        self.assertEqual(result.logits.shape, (self.model_tester.batch_size, self.model_tester.num_labels))

    def test_stablelm_sequence_classification_model_for_multi_label(self):
        config, input_dict = self.model_tester.prepare_config_and_inputs_for_common()
        config.num_labels = 3
        config.problem_type = "multi_label_classification"
        input_ids = input_dict["input_ids"]
        attention_mask = input_ids.ne(1).to(torch_device)
        sequence_labels = ids_tensor(
            [self.model_tester.batch_size, config.num_labels], self.model_tester.type_sequence_label_size
        ).to(torch.float)
        model = StableLmForSequenceClassification(config)
        model.to(torch_device)
        model.eval()
        result = model(input_ids, attention_mask=attention_mask, labels=sequence_labels)
        self.assertEqual(result.logits.shape, (self.model_tester.batch_size, self.model_tester.num_labels))

    @parameterized.expand([("linear",), ("dynamic",)])
    def test_model_rope_scaling(self, scaling_type):
        config, _ = self.model_tester.prepare_config_and_inputs_for_common()
        short_input = ids_tensor([1, 10], config.vocab_size)
        long_input = ids_tensor([1, int(config.max_position_embeddings * 1.5)], config.vocab_size)

        set_seed(42)  # Fixed seed at init time so the two models get the same random weights
        original_model = StableLmModel(config)
        original_model.to(torch_device)
        original_model.eval()
        original_short_output = original_model(short_input).last_hidden_state
        original_long_output = original_model(long_input).last_hidden_state

        set_seed(42)  # Fixed seed at init time so the two models get the same random weights
        config.rope_scaling = {"type": scaling_type, "factor": 10.0}
        scaled_model = StableLmModel(config)
        scaled_model.to(torch_device)
        scaled_model.eval()
        scaled_short_output = scaled_model(short_input).last_hidden_state
        scaled_long_output = scaled_model(long_input).last_hidden_state

        # Dynamic scaling does not change the RoPE embeddings until it receives an input longer than the original
        # maximum sequence length, so the outputs for the short input should match.
        if scaling_type == "dynamic":
            self.assertTrue(torch.allclose(original_short_output, scaled_short_output, atol=1e-5))
        else:
            self.assertFalse(torch.allclose(original_short_output, scaled_short_output, atol=1e-5))

        # The output should be different for long inputs
        self.assertFalse(torch.allclose(original_long_output, scaled_long_output, atol=1e-5))


@require_torch
class StableLmModelIntegrationTest(unittest.TestCase):
    @slow
    def test_model_stablelm_3b_4e1t_logits(self):
        input_ids = {"input_ids": torch.tensor([[510, 8588, 310, 1900, 9386]], dtype=torch.long, device=torch_device)}

        model = StableLmForCausalLM.from_pretrained("stabilityai/stablelm-3b-4e1t").to(torch_device)
        model.eval()

        output = model(**input_ids).logits

        # Expected mean on dim = -1
        EXPECTED_MEAN = torch.tensor([[2.7146, 2.4245, 1.5616, 1.4424, 2.6790]]).to(torch_device)
        self.assertTrue(torch.allclose(output.mean(dim=-1), EXPECTED_MEAN, atol=1e-4, rtol=1e-4))

        # Expected logits sliced from [0, 0, 0:30]
        EXPECTED_SLICE = torch.tensor([7.1030, -1.4195,  9.9206,  7.7008,  4.9891,  4.2169,  5.5426,  3.7878, 6.7593,  5.7360,  8.4691,  5.5448,  5.0544, 10.4129,  8.5573, 13.0405, 7.3265,  3.5868,  6.1106,  5.9406,  5.6376,  5.7490,  5.4850,  4.8124, 5.1991,  4.6419,  4.5719,  9.9588,  6.7222,  4.5070]).to(torch_device)  # fmt: skip
        self.assertTrue(torch.allclose(output[0, 0, :30], EXPECTED_SLICE, atol=1e-4, rtol=1e-4))

    @slow
    def test_model_stablelm_3b_4e1t_generation(self):
        tokenizer = AutoTokenizer.from_pretrained("stabilityai/stablelm-3b-4e1t")
        model = StableLmForCausalLM.from_pretrained("stabilityai/stablelm-3b-4e1t")
        input_ids = tokenizer.encode(
            "My favorite food has always been pizza, but lately",
            return_tensors="pt",
        )

        outputs = model.generate(input_ids, max_new_tokens=20, temperature=0)
        text = tokenizer.decode(outputs[0], skip_special_tokens=True)

        EXPECTED_TEXT_COMPLETION = """My favorite food has always been pizza, but lately I’ve been craving something different. I’ve been trying to eat healthier and I’ve"""
        self.assertEqual(text, EXPECTED_TEXT_COMPLETION)

    @require_bitsandbytes
    @slow
    @require_flash_attn
    def test_model_3b_long_prompt(self):
        EXPECTED_OUTPUT_TOKEN_IDS = [3, 3, 3]
        input_ids = [306, 338] * 2047
        model = StableLmForCausalLM.from_pretrained(
            "stabilityai/stablelm-3b-4e1t",
            device_map="auto",
            torch_dtype="auto",
            load_in_4bit=True,
            attn_implementation="flash_attention_2",
        )
        input_ids = torch.tensor([input_ids]).to(model.model.embed_tokens.weight.device)
        generated_ids = model.generate(input_ids, max_new_tokens=4, temperature=0)
        self.assertEqual(EXPECTED_OUTPUT_TOKEN_IDS, generated_ids[0][-3:].tolist())

    # Copied from transformers.tests.models.llama.test_modeling_llama.LlamaModelTest.test_eager_matches_sdpa_generate with Llama->StableLm,saibo/llama-1B->stabilityai/stablelm-3b-4e1t
    @require_torch_sdpa
    @slow
    def test_eager_matches_sdpa_generate(self):
        """
        Overwritting the common test as the test is flaky on tiny models
        """
        max_new_tokens = 30

        tokenizer = AutoTokenizer.from_pretrained("stabilityai/stablelm-3b-4e1t")

        model_sdpa = StableLmForCausalLM.from_pretrained(
            "stabilityai/stablelm-3b-4e1t",
            torch_dtype=torch.float16,
            low_cpu_mem_usage=True,
        ).to(torch_device)

        self.assertTrue(model_sdpa.config._attn_implementation == "sdpa")

        model_eager = StableLmForCausalLM.from_pretrained(
            "stabilityai/stablelm-3b-4e1t",
            torch_dtype=torch.float16,
            low_cpu_mem_usage=True,
            attn_implementation="eager",
        ).to(torch_device)

        self.assertTrue(model_eager.config._attn_implementation == "eager")

        for name, submodule in model_eager.named_modules():
            if "SdpaAttention" in submodule.__class__.__name__:
                raise ValueError("The eager model should not have SDPA attention layers")

        has_sdpa = False
        for name, submodule in model_sdpa.named_modules():
            if "SdpaAttention" in submodule.__class__.__name__:
                has_sdpa = True
                break
        if not has_sdpa:
            raise ValueError("The SDPA model should have SDPA attention layers")

        texts = [
            "hi here's a longer context, getting longer and",
            "Hello this is a very long sentence my friend, very long for real",
            "Today I am in Paris and",
        ]

        for padding_side in ["left", "right"]:
            tokenizer.padding_side = padding_side
            tokenizer.pad_token = tokenizer.eos_token

            inputs = tokenizer(texts, return_tensors="pt", padding=True).to(torch_device)

            res_eager = model_eager.generate(**inputs, max_new_tokens=max_new_tokens, do_sample=False)
            res_sdpa = model_sdpa.generate(**inputs, max_new_tokens=max_new_tokens, do_sample=False)

            with self.subTest(f"{padding_side}"):
                torch.testing.assert_close(
                    res_eager,
                    res_sdpa,
                    msg=f"\n{tokenizer.batch_decode(res_eager)} \nvs\n{tokenizer.batch_decode(res_sdpa)}",
                )