transformers/tests/models/cohere/test_modeling_cohere.py

# coding=utf-8
# Copyright 2022 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 Cohere model."""

import unittest

from transformers import CohereConfig, is_torch_available
from transformers.testing_utils import (
    require_bitsandbytes,
    require_torch,
    require_torch_multi_gpu,
    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, CohereForCausalLM, CohereModel


# Copied from transformers.tests.models.llama.LlamaModelTester with Llama->Cohere
class CohereModelTester:
    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=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,
        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.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

    # Ignore copy
    def get_config(self):
        return CohereConfig(
            vocab_size=self.vocab_size,
            hidden_size=self.hidden_size,
            num_hidden_layers=self.num_hidden_layers,
            num_attention_heads=self.num_attention_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,
            eos_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 = CohereModel(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 = CohereModel(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 = CohereForCausalLM(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 = CohereForCausalLM(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
class CohereModelTest(ModelTesterMixin, GenerationTesterMixin, PipelineTesterMixin, unittest.TestCase):
    all_model_classes = (CohereModel, CohereForCausalLM) if is_torch_available() else ()
    all_generative_model_classes = (CohereForCausalLM,) if is_torch_available() else ()
    pipeline_model_mapping = (
        {
            "feature-extraction": CohereModel,
            "text-generation": CohereForCausalLM,
        }
        if is_torch_available()
        else {}
    )
    test_headmasking = False
    test_pruning = False
    fx_compatible = True

    # Need to use `0.8` instead of `0.9` for `test_cpu_offload`
    # This is because we are hitting edge cases with the causal_mask buffer
    model_split_percents = [0.5, 0.7, 0.8]

    def setUp(self):
        self.model_tester = CohereModelTester(self)
        self.config_tester = ConfigTester(self, config_class=CohereConfig, 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_model_various_embeddings(self):
        config_and_inputs = self.model_tester.prepare_config_and_inputs()
        for type in ["absolute", "relative_key", "relative_key_query"]:
            config_and_inputs[0].position_embedding_type = type
            self.model_tester.create_and_check_model(*config_and_inputs)

    @require_bitsandbytes
    @require_torch_sdpa
    @require_torch_multi_gpu
    @slow
    def test_eager_matches_sdpa_generate(self):
        """
        Overwritting the common test as the test is flaky on tiny models
        """
        max_new_tokens = 30

        model_id = "CohereForAI/c4ai-command-r-v01-4bit"
        tokenizer = AutoTokenizer.from_pretrained(model_id)

        model_sdpa = CohereForCausalLM.from_pretrained(
            model_id, torch_dtype=torch.float16, low_cpu_mem_usage=True, device_map="auto"
        )
        self.assertTrue(model_sdpa.config._attn_implementation == "sdpa")

        model_eager = CohereForCausalLM.from_pretrained(
            model_id, torch_dtype=torch.float16, attn_implementation="eager", device_map="auto"
        )

        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)}",
                )


@require_torch
@slow
class CohereIntegrationTest(unittest.TestCase):
    @require_torch_multi_gpu
    @require_bitsandbytes
    def test_batched_4bit(self):
        model_id = "CohereForAI/c4ai-command-r-v01-4bit"

        EXPECTED_TEXT = [
            'Hello today I am going to show you how to make a simple and easy card using the new stamp set called "Hello" from the Occasions catalog. This set is so versatile and can be used for many occasions. I used the new In',
            "Hi there, here we are again with another great collection of free fonts. This time we have gathered 10 free fonts that you can download and use in your designs. These fonts are free for personal and commercial use. So",
        ]

        model = CohereForCausalLM.from_pretrained(model_id)
        tokenizer = AutoTokenizer.from_pretrained(model_id)

        tokenizer.pad_token = tokenizer.eos_token

        text = ["Hello today I am going to show you how to", "Hi there, here we are"]
        inputs = tokenizer(text, return_tensors="pt", padding=True).to(torch_device)

        output = model.generate(**inputs, max_new_tokens=40, do_sample=False)
        self.assertEqual(tokenizer.batch_decode(output, skip_special_tokens=True), EXPECTED_TEXT)

    @require_torch_sdpa
    def test_batched_small_model_logits(self):
        # Since the model is very large, we created a random cohere model so that we can do a simple
        # logits check on it.
        model_id = "hf-internal-testing/cohere-random"

        EXPECTED_LOGITS = torch.Tensor(
            [
                [[0.0000, 0.1866, -0.1997], [0.0000, -0.0736, 0.1785], [0.0000, -0.1965, -0.0569]],
                [[0.0000, -0.0302, 0.1488], [0.0000, -0.0402, 0.1351], [0.0000, -0.0341, 0.1116]],
            ]
        ).to(torch_device)

        tokenizer = AutoTokenizer.from_pretrained(model_id)
        model = CohereForCausalLM.from_pretrained(model_id, low_cpu_mem_usage=True, torch_dtype=torch.float16).to(
            torch_device
        )

        tokenizer.pad_token = tokenizer.eos_token

        text = ["Hello today I am going to show you how to", "Hi there, here we are"]
        inputs = tokenizer(text, return_tensors="pt", padding=True).to(torch_device)

        with torch.no_grad():
            output = model(**inputs)

        logits = output.logits
        self.assertTrue(torch.allclose(EXPECTED_LOGITS, logits[:, :3, :3], rtol=1e-3, atol=1e-3))