[Port] TensorFlow implementation of Mistral (#29708)

* chore: initial commit

* chore: adding imports and inits

* chore: adding the causal and classification code

* chore: adding names to the layers

* chore: using single self attn layer

* chore: built the model and layers

* chore: start with testing

* chore: docstring change, transpose fix

* fix: rotary embedding

* chore: adding cache implementation

* remove unused torch

* chore: fixing the indexing issue

* make fix-copies

* Use modeling_tf_utils.keras

* make fixup

* chore: fixing tests

* chore: adding past key value logic

* chore: adding multi label classfication test

* fix: switching on the built parameters in the layers

* fixing repo consistency

* ruff formats

* style changes

* fix: tf and pt equivalence

* removing returns from docstrings

* fix docstrings

* fix docstrings

* removing todos

* fix copies

* fix docstring

* fix docstring

* chore: using easier rotate_half

* adding integration tests

* chore: addressing review related to rotary embedding layer

* review changes

* [run-slow] mistral

* skip: test save load after resize token embedding

* style

---------

Co-authored-by: Matt <rocketknight1@gmail.com>

docs/source/en/index.md

@@ -200,7 +200,7 @@ Flax), PyTorch, and/or TensorFlow.
 |                 [Megatron-BERT](model_doc/megatron-bert)                 |       ✅        |         ❌         |      ❌      |
 |                 [Megatron-GPT2](model_doc/megatron_gpt2)                 |       ✅        |         ✅         |      ✅      |
 |                       [MGP-STR](model_doc/mgp-str)                       |       ✅        |         ❌         |      ❌      |
-|                       [Mistral](model_doc/mistral)                       |       ✅        |         ❌         |      ✅      |
+|                       [Mistral](model_doc/mistral)                       |       ✅        |         ✅         |      ✅      |
 |                       [Mixtral](model_doc/mixtral)                       |       ✅        |         ❌         |      ❌      |
 |                         [mLUKE](model_doc/mluke)                         |       ✅        |         ❌         |      ❌      |
 |                           [MMS](model_doc/mms)                           |       ✅        |         ✅         |      ✅      |

docs/source/en/model_doc/mistral.md

@@ -217,3 +217,18 @@ A list of official Hugging Face and community (indicated by 🌎) resources to h
 
 [[autodoc]] FlaxMistralForCausalLM
     - __call__
+
+## TFMistralModel
+
+[[autodoc]] TFMistralModel
+    - call
+
+## TFMistralForCausalLM
+
+[[autodoc]] TFMistralForCausalLM
+    - call
+
+## TFMistralForSequenceClassification
+
+[[autodoc]] TFMistralForSequenceClassification
+    - call

src/transformers/__init__.py

@@ -3974,6 +3974,9 @@ else:
     _import_structure["models.mbart"].extend(
         ["TFMBartForConditionalGeneration", "TFMBartModel", "TFMBartPreTrainedModel"]
     )
+    _import_structure["models.mistral"].extend(
+        ["TFMistralForCausalLM", "TFMistralForSequenceClassification", "TFMistralModel", "TFMistralPreTrainedModel"]
+    )
     _import_structure["models.mobilebert"].extend(
         [
             "TFMobileBertForMaskedLM",
@@ -8067,6 +8070,12 @@ if TYPE_CHECKING:
             TFMBartModel,
             TFMBartPreTrainedModel,
         )
+        from .models.mistral import (
+            TFMistralForCausalLM,
+            TFMistralForSequenceClassification,
+            TFMistralModel,
+            TFMistralPreTrainedModel,
+        )
         from .models.mobilebert import (
             TFMobileBertForMaskedLM,
             TFMobileBertForMultipleChoice,

src/transformers/models/auto/modeling_tf_auto.py

@@ -65,6 +65,7 @@ TF_MODEL_MAPPING_NAMES = OrderedDict(
         ("lxmert", "TFLxmertModel"),
         ("marian", "TFMarianModel"),
         ("mbart", "TFMBartModel"),
+        ("mistral", "TFMistralModel"),
         ("mobilebert", "TFMobileBertModel"),
         ("mobilevit", "TFMobileViTModel"),
         ("mpnet", "TFMPNetModel"),
@@ -178,6 +179,7 @@ TF_MODEL_FOR_CAUSAL_LM_MAPPING_NAMES = OrderedDict(
         ("gpt-sw3", "TFGPT2LMHeadModel"),
         ("gpt2", "TFGPT2LMHeadModel"),
         ("gptj", "TFGPTJForCausalLM"),
+        ("mistral", "TFMistralForCausalLM"),
         ("openai-gpt", "TFOpenAIGPTLMHeadModel"),
         ("opt", "TFOPTForCausalLM"),
         ("rembert", "TFRemBertForCausalLM"),
@@ -320,6 +322,7 @@ TF_MODEL_FOR_SEQUENCE_CLASSIFICATION_MAPPING_NAMES = OrderedDict(
         ("layoutlm", "TFLayoutLMForSequenceClassification"),
         ("layoutlmv3", "TFLayoutLMv3ForSequenceClassification"),
         ("longformer", "TFLongformerForSequenceClassification"),
+        ("mistral", "TFMistralForSequenceClassification"),
         ("mobilebert", "TFMobileBertForSequenceClassification"),
         ("mpnet", "TFMPNetForSequenceClassification"),
         ("openai-gpt", "TFOpenAIGPTForSequenceClassification"),

src/transformers/models/mistral/__init__.py

@@ -13,7 +13,13 @@
 # limitations under the License.
 from typing import TYPE_CHECKING
 
-from ...utils import OptionalDependencyNotAvailable, _LazyModule, is_flax_available, is_torch_available
+from ...utils import (
+    OptionalDependencyNotAvailable,
+    _LazyModule,
+    is_flax_available,
+    is_tf_available,
+    is_torch_available,
+)
 
 
 _import_structure = {
@@ -47,6 +53,19 @@ else:
         "FlaxMistralPreTrainedModel",
     ]
 
+try:
+    if not is_tf_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
+    _import_structure["modeling_tf_mistral"] = [
+        "TFMistralModel",
+        "TFMistralForCausalLM",
+        "TFMistralForSequenceClassification",
+        "TFMistralPreTrainedModel",
+    ]
+
 
 if TYPE_CHECKING:
     from .configuration_mistral import MistralConfig
@@ -77,6 +96,19 @@ if TYPE_CHECKING:
             FlaxMistralPreTrainedModel,
         )
 
+    try:
+        if not is_tf_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
+        from .modeling_tf_mistral import (
+            TFMistralForCausalLM,
+            TFMistralForSequenceClassification,
+            TFMistralModel,
+            TFMistralPreTrainedModel,
+        )
+
 
 else:
     import sys

src/transformers/utils/dummy_tf_objects.py

@@ -1801,6 +1801,34 @@ class TFMBartPreTrainedModel(metaclass=DummyObject):
         requires_backends(self, ["tf"])
 
 
+class TFMistralForCausalLM(metaclass=DummyObject):
+    _backends = ["tf"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["tf"])
+
+
+class TFMistralForSequenceClassification(metaclass=DummyObject):
+    _backends = ["tf"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["tf"])
+
+
+class TFMistralModel(metaclass=DummyObject):
+    _backends = ["tf"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["tf"])
+
+
+class TFMistralPreTrainedModel(metaclass=DummyObject):
+    _backends = ["tf"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["tf"])
+
+
 class TFMobileBertForMaskedLM(metaclass=DummyObject):
     _backends = ["tf"]
 

tests/models/mistral/test_modeling_tf_mistral.py

@@ -0,0 +1,367 @@
+# coding=utf-8
+# Copyright 2024 Mistral AI 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 TF 2.0 Mistral model."""
+
+import unittest
+
+import numpy as np
+
+from transformers import AutoTokenizer, MistralConfig, is_tf_available
+from transformers.testing_utils import (
+    require_tf,
+    slow,
+)
+
+from ...generation.test_tf_utils import TFGenerationIntegrationTests
+from ...test_configuration_common import ConfigTester
+from ...test_modeling_tf_common import TFModelTesterMixin, ids_tensor, random_attention_mask
+from ...test_pipeline_mixin import PipelineTesterMixin
+
+
+if is_tf_available():
+    import tensorflow as tf
+
+    from transformers.models.mistral.modeling_tf_mistral import (
+        TFMistralForCausalLM,
+        TFMistralForSequenceClassification,
+        TFMistralModel,
+    )
+
+
+class TFMistralModelTester:
+    def __init__(self, parent):
+        self.parent = parent
+        self.batch_size = 13
+        self.seq_length = 7
+        self.is_training = True
+        self.use_input_mask = True
+        self.use_token_type_ids = False
+        self.use_labels = True
+        self.vocab_size = 99
+        self.hidden_size = 32
+        self.num_hidden_layers = 2
+        self.num_attention_heads = 4
+        self.num_key_value_heads = 2
+        self.intermediate_size = 37
+        self.hidden_act = "gelu"
+        self.hidden_dropout_prob = 0.1
+        self.attention_probs_dropout_prob = 0.1
+        self.max_position_embeddings = 512
+        self.type_vocab_size = 16
+        self.type_sequence_label_size = 2
+        self.initializer_range = 0.02
+        self.num_labels = 3
+        self.num_choices = 4
+        self.pad_token_id = 0
+        self.scope = None
+        self.bos_token_id = self.vocab_size - 1
+        self.eos_token_id = self.vocab_size - 1
+        self.pad_token_id = self.vocab_size - 1
+
+    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 = random_attention_mask([self.batch_size, self.seq_length], self.vocab_size)
+
+        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 = MistralConfig(
+            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,
+        )
+
+        return (
+            config,
+            input_ids,
+            token_type_ids,
+            input_mask,
+            sequence_labels,
+            token_labels,
+            choice_labels,
+        )
+
+    def create_and_check_model(
+        self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
+    ):
+        model = TFMistralModel(config=config)
+        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 = TFMistralModel(config)
+        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 = TFMistralForCausalLM(config=config)
+        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 = TFMistralForCausalLM(config=config)
+
+        # 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 = tf.concat([input_ids, next_tokens], axis=-1)
+        next_attention_mask = tf.concat([input_mask, next_mask], axis=-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(np.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_tf
+class TFMistralModelTest(TFModelTesterMixin, TFGenerationIntegrationTests, PipelineTesterMixin, unittest.TestCase):
+    all_model_classes = (
+        (TFMistralModel, TFMistralForCausalLM, TFMistralForSequenceClassification) if is_tf_available() else ()
+    )
+    all_generative_model_classes = (TFMistralForCausalLM,) if is_tf_available() else ()
+    pipeline_model_mapping = (
+        {
+            "feature-extraction": TFMistralModel,
+            "text-classification": TFMistralForSequenceClassification,
+            "text-generation": TFMistralForCausalLM,
+            "zero-shot": TFMistralForSequenceClassification,
+        }
+        if is_tf_available()
+        else {}
+    )
+    test_onnx = False
+    test_pruning = False
+    test_missing_keys = False
+    test_head_masking = False
+
+    # TODO (ydshieh): Check this. See https://app.circleci.com/pipelines/github/huggingface/transformers/79245/workflows/9490ef58-79c2-410d-8f51-e3495156cf9c/jobs/1012146
+    def is_pipeline_test_to_skip(
+        self, pipeline_test_casse_name, config_class, model_architecture, tokenizer_name, processor_name
+    ):
+        return True
+
+    def setUp(self):
+        self.model_tester = TFMistralModelTester(self)
+        self.config_tester = ConfigTester(self, config_class=MistralConfig, 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)
+
+    def test_Mistral_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 = tf.not_equal(input_ids, 1)
+        sequence_labels = ids_tensor([self.model_tester.batch_size], self.model_tester.type_sequence_label_size)
+        model = TFMistralForSequenceClassification(config)
+        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_Mistral_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 = tf.not_equal(input_ids, 1)
+        sequence_labels = ids_tensor([self.model_tester.batch_size], self.model_tester.type_sequence_label_size)
+        model = TFMistralForSequenceClassification(config)
+        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_Mistral_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 = tf.not_equal(input_ids, 1)
+        sequence_labels = ids_tensor([self.model_tester.batch_size], self.model_tester.type_sequence_label_size)
+        model = TFMistralForSequenceClassification(config)
+        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))
+
+    @unittest.skip("Mistral buffers include complex numbers, which breaks this test")
+    def test_save_load_fast_init_from_base(self):
+        pass
+
+    @unittest.skip("Mistral uses GQA on all models so the KV cache is a non standard format")
+    def test_past_key_values_format(self):
+        pass
+
+    @unittest.skip("Vocab resizing is not supported")
+    def test_save_load_after_resize_token_embeddings(self):
+        pass
+
+
+@require_tf
+class TFMistralIntegrationTest(unittest.TestCase):
+    @slow
+    def test_model_7b_logits(self):
+        input_ids = [1, 306, 4658, 278, 6593, 310, 2834, 338]
+        model = TFMistralForCausalLM.from_pretrained(
+            "hf-internal-testing/tiny-random-MistralForCausalLM", from_pt=True
+        )
+        input_ids = tf.constant([input_ids])
+        out = model(input_ids).logits
+        # Expected mean on dim = -1
+        EXPECTED_MEAN = tf.constant(
+            [[-1.281e-04, -2.869e-04, -9.989e-05, -8.995e-05, 2.494e-04, -3.083e-04, -2.672e-04, -1.239e-04]]
+        )
+        tf.debugging.assert_near(tf.reduce_mean(out, axis=-1), EXPECTED_MEAN, atol=1e-2, rtol=1e-2)
+        # slicing logits[0, 0, 0:30]
+        EXPECTED_SLICE = tf.constant([0.1033,  0.1493, -0.0041, -0.0021, -0.1686,  0.0356,  0.0812,  0.2218, -0.1257,  0.1920,  0.0929,  0.1181,  0.0111,  0.0395, -0.0064,  0.1712, -0.0751,  0.0625, -0.2409,  0.1541, -0.1271, -0.2296, -0.0099, -0.0160, 0.0311, -0.0824, -0.1518,  0.0722,  0.0187,  0.0484])  # fmt: skip
+        tf.debugging.assert_near(out[0, 0, :30], EXPECTED_SLICE, atol=1e-4, rtol=1e-4)
+
+    @slow
+    def test_model_7b_generation(self):
+        EXPECTED_TEXT_COMPLETION = """My favourite condiment is  Werk a EgyadjustPrintfigiousPDFPHPct guns Ein motor conceti barSequ内 infrastructure millretval"""
+        prompt = "My favourite condiment is "
+        tokenizer = AutoTokenizer.from_pretrained("hf-internal-testing/tiny-random-MistralForCausalLM", use_fast=False)
+        model = TFMistralForCausalLM.from_pretrained(
+            "hf-internal-testing/tiny-random-MistralForCausalLM", from_pt=True
+        )
+        input_ids = tokenizer.encode(prompt, return_tensors="tf")
+
+        # greedy generation outputs
+        generated_ids = model.generate(input_ids, max_new_tokens=20, temperature=0)
+        text = tokenizer.decode(generated_ids[0], skip_special_tokens=True)
+        self.assertEqual(EXPECTED_TEXT_COMPLETION, text)