[ `gemma`] Adds support for Gemma 💎 (#29167)

* inital commit

* update

* update conversion checkpoint

* update conversion script

* nits

* some fixes

* nits

* merge

* fix permute

* nits

* fix

* nits

* nits

* nits

* fix rope

* fix both rope

* nites

* style

* make sure flax works

* fix flax init code

* fix foward

* nits

* print flax generation out

* current code

* nits

* SIIIIIIIIIIIIIIIIIII

* update

* add new tokenizer

* correct fast tokenizer

* fix conversion

* more comments

* fix modeling and conversion

* nits and nits

* nits testing

* add some tokenization tests

* add some edge cases

* add slow tests and fix them

* fixup

* fix copies for modeling

* fix copies

* add 7B slow tests

* fix

* fix

* fix tests

* make tokenizer cis go green

* styling

* last tokenizer nits

* update jax tests

* fix flax for 7b

* add jit testing 🤗



* cleanups

* isolated nit, inv_freq for rotary_emb.inv_freq

* propagate to jax

* Apply suggestions from code review
Co-authored-by: Sanchit Gandhi <93869735+sanchit-gandhi@users.noreply.github.com>

* adjust test

* fix conversion script

* change name

* correct file names

* update conversion script

* Fix bos and eos token ids in the model configuration (#3)

* update modelling

* update conversion script

* add static cache for gemma

* fix sdpa generate

* fix batched

* multiple fixes

* fix FA2

* final fix

* Rename a few missing strings and filenames (#4)

* merge with upstream main

* fix copies

* fix copies

* fix fixup

* fix fixup

* fix

* fix

* final tests

* fix fx gemma tests

* fix fx bf16/fp16 tests

* update slow fx tests

* fx slow tests: one logits, one generation

* move jit test standalone

* Apply suggestions from code review

* nits

* tokenizer updates

* more tokenization updates: custom GemmaSentencepieceExtrator

* style

* Update src/transformers/cache_utils.py

* Update src/transformers/models/gemma/__init__.py

* Update tests/models/gemma/test_modeling_flax_gemma.py

* small nits

* style

* update tokenization test

* fix the rotary embedding

* with style

* fix slow tests

* WARNING this commit might be very important for precisions

* Update tests/models/gemma/test_modeling_flax_gemma.py

* Update src/transformers/models/gemma/configuration_gemma.py
Co-authored-by: Lysandre Debut <hi@lysand.re>

* Update src/transformers/models/gemma/modeling_flax_gemma.py
Co-authored-by: Lysandre Debut <hi@lysand.re>

* small nits here and there!

* forgotten nit

* remove on the fly computation of inv_freq

* revert previous change, let's be safe and for now re-compute freq cis to make sure it's in float

* Apply suggestions from code review
Co-authored-by: Pedro Cuenca <pedro@huggingface.co>

* Update src/transformers/models/gemma/convert_gemma_weights_to_hf.py
Co-authored-by: Pedro Cuenca <pedro@huggingface.co>

* Update src/transformers/models/gemma/convert_gemma_weights_to_hf.py
Co-authored-by: Pedro Cuenca <pedro@huggingface.co>

* Update tests/models/gemma/test_modeling_gemma.py
Co-authored-by: Pedro Cuenca <pedro@huggingface.co>

* Update tests/models/gemma/test_modeling_gemma.py
Co-authored-by: Pedro Cuenca <pedro@huggingface.co>

* Update tests/models/gemma/test_modeling_gemma.py
Co-authored-by: Pedro Cuenca <pedro@huggingface.co>

* Update tests/models/gemma/test_modeling_flax_gemma.py
Co-authored-by: Pedro Cuenca <pedro@huggingface.co>

* Update tests/models/gemma/test_modeling_gemma.py
Co-authored-by: Pedro Cuenca <pedro@huggingface.co>

* Update tests/models/gemma/test_modeling_gemma.py
Co-authored-by: Pedro Cuenca <pedro@huggingface.co>

* Update tests/models/gemma/test_tokenization_gemma.py
Co-authored-by: Pedro Cuenca <pedro@huggingface.co>

* Update tests/models/gemma/test_tokenization_gemma.py
Co-authored-by: Pedro Cuenca <pedro@huggingface.co>

* Update tests/models/gemma/test_tokenization_gemma.py
Co-authored-by: Pedro Cuenca <pedro@huggingface.co>

* Update tests/models/gemma/test_tokenization_gemma.py
Co-authored-by: Pedro Cuenca <pedro@huggingface.co>

* Update tests/models/gemma/test_modeling_gemma.py
Co-authored-by: Pedro Cuenca <pedro@huggingface.co>

* Update tests/models/gemma/test_modeling_gemma.py
Co-authored-by: Pedro Cuenca <pedro@huggingface.co>

* Update tests/models/gemma/test_modeling_gemma.py
Co-authored-by: Pedro Cuenca <pedro@huggingface.co>

* Update tests/models/gemma/test_modeling_gemma.py
Co-authored-by: Pedro Cuenca <pedro@huggingface.co>

* Update tests/models/gemma/test_modeling_gemma.py
Co-authored-by: Pedro Cuenca <pedro@huggingface.co>

* nit conversion script link

* fix some tests

* add not doctest and pr doctest

* repo consistency

* fix last CIs 🚀



* update all readmes

---------
Co-authored-by: younesbelkada <younesbelkada@gmail.com>
Co-authored-by: Sanchit Gandhi <93869735+sanchit-gandhi@users.noreply.github.com>
Co-authored-by: Pedro Cuenca <pedro@huggingface.co>
Co-authored-by: Younes Belkada <49240599+younesbelkada@users.noreply.github.com>
Co-authored-by: sanchit-gandhi <sanchit@huggingface.co>
Co-authored-by: Lysandre Debut <hi@lysand.re>

src/transformers/models/auto/modeling_auto.py

@@ -103,6 +103,7 @@ MODEL_MAPPING_NAMES = OrderedDict(
         ("focalnet", "FocalNetModel"),
         ("fsmt", "FSMTModel"),
         ("funnel", ("FunnelModel", "FunnelBaseModel")),
+        ("gemma", "GemmaModel"),
         ("git", "GitModel"),
         ("glpn", "GLPNModel"),
         ("gpt-sw3", "GPT2Model"),
@@ -426,6 +427,7 @@ MODEL_FOR_CAUSAL_LM_MAPPING_NAMES = OrderedDict(
         ("ernie", "ErnieForCausalLM"),
         ("falcon", "FalconForCausalLM"),
         ("fuyu", "FuyuForCausalLM"),
+        ("gemma", "GemmaForCausalLM"),
         ("git", "GitForCausalLM"),
         ("gpt-sw3", "GPT2LMHeadModel"),
         ("gpt2", "GPT2LMHeadModel"),
@@ -764,6 +766,7 @@ MODEL_FOR_SEQUENCE_CLASSIFICATION_MAPPING_NAMES = OrderedDict(
         ("flaubert", "FlaubertForSequenceClassification"),
         ("fnet", "FNetForSequenceClassification"),
         ("funnel", "FunnelForSequenceClassification"),
+        ("gemma", "GemmaForSequenceClassification"),
         ("gpt-sw3", "GPT2ForSequenceClassification"),
         ("gpt2", "GPT2ForSequenceClassification"),
         ("gpt_bigcode", "GPTBigCodeForSequenceClassification"),

src/transformers/models/auto/modeling_flax_auto.py

@@ -39,6 +39,7 @@ FLAX_MODEL_MAPPING_NAMES = OrderedDict(
         ("clip", "FlaxCLIPModel"),
         ("distilbert", "FlaxDistilBertModel"),
         ("electra", "FlaxElectraModel"),
+        ("gemma", "FlaxGemmaModel"),
         ("gpt-sw3", "FlaxGPT2Model"),
         ("gpt2", "FlaxGPT2Model"),
         ("gpt_neo", "FlaxGPTNeoModel"),
@@ -144,6 +145,7 @@ FLAX_MODEL_FOR_CAUSAL_LM_MAPPING_NAMES = OrderedDict(
         ("big_bird", "FlaxBigBirdForCausalLM"),
         ("bloom", "FlaxBloomForCausalLM"),
         ("electra", "FlaxElectraForCausalLM"),
+        ("gemma", "FlaxGemmaForCausalLM"),
         ("gpt-sw3", "FlaxGPT2LMHeadModel"),
         ("gpt2", "FlaxGPT2LMHeadModel"),
         ("gpt_neo", "FlaxGPTNeoForCausalLM"),

src/transformers/models/auto/tokenization_auto.py

@@ -178,6 +178,13 @@ else:
             ("fnet", ("FNetTokenizer", "FNetTokenizerFast" if is_tokenizers_available() else None)),
             ("fsmt", ("FSMTTokenizer", None)),
             ("funnel", ("FunnelTokenizer", "FunnelTokenizerFast" if is_tokenizers_available() else None)),
+            (
+                "gemma",
+                (
+                    "GemmaTokenizer" if is_sentencepiece_available() else None,
+                    "GemmaTokenizerFast" if is_tokenizers_available() else None,
+                ),
+            ),
             ("git", ("BertTokenizer", "BertTokenizerFast" if is_tokenizers_available() else None)),
             ("gpt-sw3", ("GPTSw3Tokenizer" if is_sentencepiece_available() else None, None)),
             ("gpt2", ("GPT2Tokenizer", "GPT2TokenizerFast" if is_tokenizers_available() else None)),

src/transformers/models/gemma/__init__.py

+# 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.
+from typing import TYPE_CHECKING
+
+from ...utils import (
+    OptionalDependencyNotAvailable,
+    _LazyModule,
+    is_flax_available,
+    is_sentencepiece_available,
+    is_tokenizers_available,
+    is_torch_available,
+)
+
+
+_import_structure = {
+    "configuration_gemma": ["GEMMA_PRETRAINED_CONFIG_ARCHIVE_MAP", "GemmaConfig"],
+}
+
+try:
+    if not is_sentencepiece_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
+    _import_structure["tokenization_gemma"] = ["GemmaTokenizer"]
+
+try:
+    if not is_tokenizers_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
+    _import_structure["tokenization_gemma_fast"] = ["GemmaTokenizerFast"]
+
+
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
+    _import_structure["modeling_gemma"] = [
+        "GemmaForCausalLM",
+        "GemmaModel",
+        "GemmaPreTrainedModel",
+        "GemmaForSequenceClassification",
+    ]
+
+try:
+    if not is_flax_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
+    _import_structure["modeling_flax_gemma"] = [
+        "FlaxGemmaForCausalLM",
+        "FlaxGemmaModel",
+        "FlaxGemmaPreTrainedModel",
+    ]
+
+
+if TYPE_CHECKING:
+    from .configuration_gemma import GEMMA_PRETRAINED_CONFIG_ARCHIVE_MAP, GemmaConfig
+
+    try:
+        if not is_sentencepiece_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
+        from .tokenization_gemma import GemmaTokenizer
+
+    try:
+        if not is_tokenizers_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
+        from .tokenization_gemma_fast import GemmaTokenizerFast
+
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
+        from .modeling_gemma import (
+            GemmaForCausalLM,
+            GemmaForSequenceClassification,
+            GemmaModel,
+            GemmaPreTrainedModel,
+        )
+
+    try:
+        if not is_flax_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
+        from .modeling_flax_gemma import (
+            FlaxGemmaForCausalLM,
+            FlaxGemmaModel,
+            FlaxGemmaPreTrainedModel,
+        )
+
+
+else:
+    import sys
+
+    sys.modules[__name__] = _LazyModule(__name__, globals()["__file__"], _import_structure, module_spec=__spec__)

src/transformers/models/gemma/configuration_gemma.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.
+""" Gemma model configuration"""
+
+from ...configuration_utils import PretrainedConfig
+from ...utils import logging
+
+
+logger = logging.get_logger(__name__)
+
+GEMMA_PRETRAINED_CONFIG_ARCHIVE_MAP = {}
+
+
+class GemmaConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`GemmaModel`]. It is used to instantiate an Gemma
+    model according to the specified arguments, defining the model architecture. Instantiating a configuration with the
+    defaults will yield a similar configuration to that of the Gemma-7B.
+
+    e.g. [google/gemma-7b](https://huggingface.co/google/gemma-7b)
+
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+
+    Args:
+        vocab_size (`int`, *optional*, defaults to 256000):
+            Vocabulary size of the Gemma model. Defines the number of different tokens that can be represented by the
+            `inputs_ids` passed when calling [`GemmaModel`]
+        hidden_size (`int`, *optional*, defaults to 3072):
+            Dimension of the hidden representations.
+        intermediate_size (`int`, *optional*, defaults to 24576):
+            Dimension of the MLP representations.
+        num_hidden_layers (`int`, *optional*, defaults to 28):
+            Number of hidden layers in the Transformer decoder.
+        num_attention_heads (`int`, *optional*, defaults to 16):
+            Number of attention heads for each attention layer in the Transformer decoder.
+        num_key_value_heads (`int`, *optional*, defaults to 16):
+            This is the number of key_value heads that should be used to implement Grouped Query Attention. If
+            `num_key_value_heads=num_attention_heads`, the model will use Multi Head Attention (MHA), if
+            `num_key_value_heads=1 the model will use Multi Query Attention (MQA) otherwise GQA is used. When
+            converting a multi-head checkpoint to a GQA checkpoint, each group key and value head should be constructed
+            by meanpooling all the original heads within that group. For more details checkout [this
+            paper](https://arxiv.org/pdf/2305.13245.pdf). If it is not specified, will default to
+            `num_attention_heads`.
+        head_dim (`int`, *optional*, defaults to 256):
+            The attention head dimension.
+        hidden_act (`str` or `function`, *optional*, defaults to `"gelu"`):
+            The non-linear activation function (function or string) in the decoder.
+        max_position_embeddings (`int`, *optional*, defaults to 8192):
+            The maximum sequence length that this model might ever be used with.
+        initializer_range (`float`, *optional*, defaults to 0.02):
+            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+        rms_norm_eps (`float`, *optional*, defaults to 1e-06):
+            The epsilon used by the rms normalization layers.
+        use_cache (`bool`, *optional*, defaults to `True`):
+            Whether or not the model should return the last key/values attentions (not used by all models). Only
+            relevant if `config.is_decoder=True`.
+        pad_token_id (`int`, *optional*, defaults to 0):
+            Padding token id.
+        eos_token_id (`int`, *optional*, defaults to 1):
+            End of stream token id.
+        bos_token_id (`int`, *optional*, defaults to 2):
+            Beginning of stream token id.
+        tie_word_embeddings (`bool`, *optional*, defaults to `True`):
+            Whether to tie weight embeddings
+        rope_theta (`float`, *optional*, defaults to 10000.0):
+            The base period of the RoPE embeddings.
+        attention_bias (`bool`, defaults to `False`, *optional*, defaults to `False`):
+            Whether to use a bias in the query, key, value and output projection layers during self-attention.
+        attention_dropout (`float`, *optional*, defaults to 0.0):
+            The dropout ratio for the attention probabilities.
+
+    ```python
+    >>> from transformers import GemmaModel, GemmaConfig
+
+    >>> # Initializing a Gemma gemma-7b style configuration
+    >>> configuration = GemmaConfig()
+
+    >>> # Initializing a model from the gemma-7b style configuration
+    >>> model = GemmaModel(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+
+    model_type = "gemma"
+    keys_to_ignore_at_inference = ["past_key_values"]
+
+    def __init__(
+        self,
+        vocab_size=256000,
+        hidden_size=3072,
+        intermediate_size=24576,
+        num_hidden_layers=28,
+        num_attention_heads=16,
+        num_key_value_heads=16,
+        head_dim=256,
+        hidden_act="gelu",
+        max_position_embeddings=8192,
+        initializer_range=0.02,
+        rms_norm_eps=1e-6,
+        use_cache=True,
+        pad_token_id=0,
+        eos_token_id=1,
+        bos_token_id=2,
+        tie_word_embeddings=True,
+        rope_theta=10000.0,
+        attention_bias=False,
+        attention_dropout=0.0,
+        **kwargs,
+    ):
+        self.vocab_size = vocab_size
+        self.max_position_embeddings = max_position_embeddings
+        self.hidden_size = hidden_size
+        self.intermediate_size = intermediate_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.head_dim = head_dim
+        self.num_key_value_heads = num_key_value_heads
+        self.hidden_act = hidden_act
+        self.initializer_range = initializer_range
+        self.rms_norm_eps = rms_norm_eps
+        self.use_cache = use_cache
+        self.rope_theta = rope_theta
+        self.attention_bias = attention_bias
+        self.attention_dropout = attention_dropout
+
+        super().__init__(
+            pad_token_id=pad_token_id,
+            bos_token_id=bos_token_id,
+            eos_token_id=eos_token_id,
+            tie_word_embeddings=tie_word_embeddings,
+            **kwargs,
+        )

src/transformers/models/gemma/convert_gemma_weights_to_hf.py

+# 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.
+import argparse
+import os
+import warnings
+
+import torch
+from accelerate import init_empty_weights
+
+from transformers import GemmaConfig, GemmaForCausalLM, GemmaTokenizer
+
+
+try:
+    from transformers import GemmaTokenizerFast
+except ImportError as e:
+    warnings.warn(e)
+    warnings.warn(
+        "The converted tokenizer will be the `slow` tokenizer. To use the fast, update your `tokenizers` library and re-run the tokenizer conversion"
+    )
+    GemmaTokenizerFast = None
+
+"""
+Sample usage:
+
+```
+python src/transformers/models/gemma/convert_gemma_weights_to_hf.py \
+    --input_dir /path/to/downloaded/gemma/weights --model_size 7B --output_dir /output/path
+```
+
+Thereafter, models can be loaded via:
+
+```py
+from transformers import GemmaForCausalLM, GemmaTokenizerFast
+
+model = GemmaForCausalLM.from_pretrained("/output/path")
+tokenizer = GemmaTokenizerFast.from_pretrained("/output/path")
+```
+
+Important note: you need to be able to host the whole model in RAM to execute this script (even if the biggest versions
+come in several checkpoints they each contain a part of each weight of the model, so we need to load them all in RAM).
+"""
+
+gemma_2b_config = GemmaConfig(
+    num_hidden_layers=18,
+    num_attention_heads=8,
+    num_key_value_heads=1,
+    hidden_size=2048,
+    intermediate_size=16384,
+)
+
+gemma_7b_config = GemmaConfig()
+
+CONFIG_MAPPING = {"2B": gemma_2b_config, "7B": gemma_7b_config}
+LAYER_NAME_MAPPING = {"embedder.weight": "model.embed_tokens.weight"}
+
+
+def write_model(save_path, input_base_path, config, safe_serialization=True, push_to_hub=False):
+    num_attn_heads = config.num_attention_heads
+    hidden_size = config.hidden_size
+    num_kv_heads = config.num_key_value_heads
+    head_dim = config.head_dim
+
+    print(f"Fetching all parameters from the checkpoint at '{input_base_path}'")
+    model_state_dict = torch.load(input_base_path, map_location="cpu")["model_state_dict"]
+    model_state_dict.pop("freqs_cis")
+
+    state_dict = {}
+    for k, v in model_state_dict.items():
+        if "qkv_proj" in k:
+            if num_kv_heads == 1:
+                v = v.reshape(num_attn_heads + num_kv_heads * 2, head_dim, hidden_size)
+                q_proj = v[:num_attn_heads, ...]
+                k_proj = v[num_attn_heads : num_attn_heads + num_kv_heads, ...].repeat(num_kv_heads, 1, 1)
+                v_proj = v[-num_kv_heads:, ...].repeat(num_kv_heads, 1, 1)
+
+                state_dict[k.replace("qkv_proj", "q_proj")] = q_proj.reshape(
+                    num_attn_heads * head_dim, hidden_size
+                ).clone()
+                state_dict[k.replace("qkv_proj", "k_proj")] = k_proj.reshape(
+                    num_kv_heads * head_dim, hidden_size
+                ).clone()
+                state_dict[k.replace("qkv_proj", "v_proj")] = v_proj[0].clone()
+            else:
+                q_proj, k_proj, v_proj = torch.split(v, v.shape[0] // 3, 0)
+                state_dict[k.replace("qkv_proj", "q_proj")] = q_proj.reshape(
+                    num_attn_heads * head_dim, hidden_size
+                ).clone()
+                state_dict[k.replace("qkv_proj", "k_proj")] = k_proj.reshape(
+                    num_kv_heads * head_dim, hidden_size
+                ).clone()
+                state_dict[k.replace("qkv_proj", "v_proj")] = v_proj.clone()
+
+        elif k == "embedder.weight":
+            state_dict[LAYER_NAME_MAPPING[k]] = v
+            state_dict["lm_head.weight"] = v
+        else:
+            state_dict[k] = v
+
+    print("Loading the checkpoint in a Gemma model.")
+    with init_empty_weights():
+        model = GemmaForCausalLM(config)
+    model.load_state_dict(state_dict, assign=True, strict=False)
+
+    model.config.torch_dtype = torch.float32
+    del model.config._name_or_path
+    print("Saving in the Transformers format.")
+
+    if push_to_hub:
+        print(f"pushing the model to {save_path}")
+        model.push_to_hub(save_path, safe_serialization=safe_serialization, private=True)
+    else:
+        model.save_pretrained(save_path, safe_serialization=safe_serialization)
+
+
+def write_tokenizer(input_tokenizer_path, save_path, push_to_hub=False):
+    # Initialize the tokenizer based on the `spm` model
+    tokenizer_class = GemmaTokenizer if GemmaTokenizerFast is None else GemmaTokenizerFast
+    print(f"Saving a {tokenizer_class.__name__} to {save_path}.")
+    tokenizer = tokenizer_class(input_tokenizer_path)
+    if push_to_hub:
+        tokenizer.push_to_hub(save_path)
+    else:
+        tokenizer.save_pretrained(save_path)
+
+
+def main():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--input_checkpoint",
+        help="Absolute path to the target Gemma weights.",
+        required=True,
+    )
+    parser.add_argument(
+        "--tokenizer_checkpoint",
+        help="Location of Gemma tokenizer model",
+    )
+    parser.add_argument(
+        "--model_size",
+        default="7B",
+        choices=["2B", "7B", "tokenizer_only"],
+        help="'f' models correspond to the finetuned versions, and are specific to the Gemma2 official release. For more details on Gemma2, checkout the original repo: https://huggingface.co/google/gemma-7b",
+    )
+    parser.add_argument(
+        "--output_dir",
+        default="google/gemma-7b",
+        help="Location to write HF model and tokenizer",
+    )
+    parser.add_argument(
+        "--pickle_serialization",
+        help="Whether or not to save using `safetensors`.",
+        action="store_true",
+        default=False,
+    )
+    parser.add_argument(
+        "--convert_tokenizer",
+        help="Whether or not to convert the tokenizer as well.",
+        action="store_true",
+        default=False,
+    )
+    parser.add_argument(
+        "--push_to_hub",
+        help="Whether or not to push the model to the hub at `output_dir` instead of saving it locally.",
+        action="store_true",
+        default=False,
+    )
+    args = parser.parse_args()
+
+    if args.convert_tokenizer:
+        if args.tokenizer_checkpoint is None:
+            raise ValueError("Path to the tokenizer is required when passing --convert_tokenizer")
+
+        spm_path = os.path.join(args.tokenizer_checkpoint)
+        write_tokenizer(spm_path, args.output_dir, args.push_to_hub)
+
+    config = CONFIG_MAPPING[args.model_size]
+    write_model(
+        config=config,
+        input_base_path=args.input_checkpoint,
+        save_path=args.output_dir,
+        safe_serialization=not args.pickle_serialization,
+        push_to_hub=args.push_to_hub,
+    )
+
+
+if __name__ == "__main__":
+    main()

src/transformers/models/gemma/tokenization_gemma.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.
+
+"""Tokenization classes for Gemma."""
+import os
+from shutil import copyfile
+from typing import TYPE_CHECKING, Any, Dict, List, Optional, Tuple
+
+import sentencepiece as spm
+
+from ...tokenization_utils import AddedToken, PreTrainedTokenizer
+from ...utils import logging
+
+
+if TYPE_CHECKING:
+    pass
+
+logger = logging.get_logger(__name__)
+
+VOCAB_FILES_NAMES = {"vocab_file": "tokenizer.model"}
+
+SPIECE_UNDERLINE = "▁"
+
+
+class GemmaTokenizer(PreTrainedTokenizer):
+    """
+    Construct a Gemma tokenizer. Based on byte-level Byte-Pair-Encoding. The default padding token is unset as there is
+    no padding token in the original model.
+
+    Args:
+        vocab_file (`str`):
+            Path to the vocabulary file.
+        unk_token (`str` or `tokenizers.AddedToken`, *optional*, defaults to `"<unk>"`):
+            The unknown token. A token that is not in the vocabulary cannot be converted to an ID and is set to be this
+            token instead.
+        bos_token (`str` or `tokenizers.AddedToken`, *optional*, defaults to `"<bos>"`):
+            The beginning of sequence token that was used during pretraining. Can be used a sequence classifier token.
+        eos_token (`str` or `tokenizers.AddedToken`, *optional*, defaults to `"<eos>"`):
+            The end of sequence token.
+        pad_token (`str` or `tokenizers.AddedToken`, *optional*, defaults to `"<pad>"`):
+            A special token used to make arrays of tokens the same size for batching purpose. Will then be ignored by
+            attention mechanisms or loss computation.
+        sp_model_kwargs (`Dict[str, Any]`, `Optional`, *optional*):
+            Will be passed to the `SentencePieceProcessor.__init__()` method. The [Python wrapper for
+            SentencePiece](https://github.com/google/sentencepiece/tree/master/python) can be used, among other things,
+            to set:
+
+            - `enable_sampling`: Enable subword regularization.
+            - `nbest_size`: Sampling parameters for unigram. Invalid for BPE-Dropout.
+
+              - `nbest_size = {0,1}`: No sampling is performed.
+              - `nbest_size > 1`: samples from the nbest_size results.
+              - `nbest_size < 0`: assuming that nbest_size is infinite and samples from the all hypothesis (lattice)
+                using forward-filtering-and-backward-sampling algorithm.
+
+            - `alpha`: Smoothing parameter for unigram sampling, and dropout probability of merge operations for
+              BPE-dropout.
+
+        add_bos_token (`bool`, *optional*, defaults to `True`):
+            Whether or not to add an `bos_token` at the start of sequences.
+        add_eos_token (`bool`, *optional*, defaults to `False`):
+            Whether or not to add an `eos_token` at the end of sequences.
+        clean_up_tokenization_spaces (`bool`, *optional*, defaults to `False`):
+            Whether or not to cleanup spaces after decoding, cleanup consists in removing potential artifacts like
+            extra spaces.
+        use_default_system_prompt (`bool`, *optional*, defaults to `False`):
+            Whether or not the default system prompt for Gemma should be used.
+        spaces_between_special_tokens (`bool`, *optional*, defaults to `False`):
+            Whether or not to add spaces between special tokens.
+    """
+
+    vocab_files_names = VOCAB_FILES_NAMES
+    model_input_names = ["input_ids", "attention_mask"]
+
+    def __init__(
+        self,
+        vocab_file,
+        unk_token="<unk>",
+        bos_token="<bos>",
+        eos_token="<eos>",
+        pad_token="<pad>",
+        sp_model_kwargs: Optional[Dict[str, Any]] = None,
+        add_bos_token=True,
+        add_eos_token=False,
+        clean_up_tokenization_spaces=False,
+        use_default_system_prompt=False,
+        spaces_between_special_tokens=False,
+        **kwargs,
+    ):
+        self.sp_model_kwargs = {} if sp_model_kwargs is None else sp_model_kwargs
+        bos_token = AddedToken(bos_token, normalized=False, special=True) if isinstance(bos_token, str) else bos_token
+        eos_token = AddedToken(eos_token, normalized=False, special=True) if isinstance(eos_token, str) else eos_token
+        unk_token = AddedToken(unk_token, normalized=False, special=True) if isinstance(unk_token, str) else unk_token
+        pad_token = AddedToken(pad_token, normalized=False, special=True) if isinstance(pad_token, str) else pad_token
+
+        self.vocab_file = vocab_file
+        self.add_bos_token = add_bos_token
+        self.add_eos_token = add_eos_token
+        self.use_default_system_prompt = use_default_system_prompt
+
+        self.sp_model = spm.SentencePieceProcessor(**self.sp_model_kwargs)
+        self.sp_model.Load(vocab_file)
+
+        super().__init__(
+            bos_token=bos_token,
+            eos_token=eos_token,
+            unk_token=unk_token,
+            pad_token=pad_token,
+            add_bos_token=add_bos_token,
+            add_eos_token=add_eos_token,
+            sp_model_kwargs=self.sp_model_kwargs,
+            clean_up_tokenization_spaces=clean_up_tokenization_spaces,
+            use_default_system_prompt=use_default_system_prompt,
+            spaces_between_special_tokens=spaces_between_special_tokens,
+            **kwargs,
+        )
+
+    # Copied from transformers.models.llama.tokenization_llama.LlamaTokenizer.__getstate__
+    def __getstate__(self):
+        state = self.__dict__.copy()
+        state["sp_model"] = None
+        state["sp_model_proto"] = self.sp_model.serialized_model_proto()
+        return state
+
+    # Copied from transformers.models.llama.tokenization_llama.LlamaTokenizer.__setstate__
+    def __setstate__(self, d):
+        self.__dict__ = d
+        self.sp_model = spm.SentencePieceProcessor(**self.sp_model_kwargs)
+        self.sp_model.LoadFromSerializedProto(self.sp_model_proto)
+
+    @property
+    # Copied from transformers.models.llama.tokenization_llama.LlamaTokenizer.vocab_size
+    def vocab_size(self):
+        """Returns vocab size"""
+        return self.sp_model.get_piece_size()
+
+    # Copied from transformers.models.llama.tokenization_llama.LlamaTokenizer.get_vocab
+    def get_vocab(self):
+        """Returns vocab as a dict"""
+        vocab = {self.convert_ids_to_tokens(i): i for i in range(self.vocab_size)}
+        vocab.update(self.added_tokens_encoder)
+        return vocab
+
+    def _tokenize(self, text, **kwargs):
+        """
+        Returns a tokenized string. The Gemma tokenizer never adds a prefix space.
+        """
+        return self.sp_model.encode(text, out_type=str)
+
+    # Copied from transformers.models.llama.tokenization_llama.LlamaTokenizer._convert_token_to_id
+    def _convert_token_to_id(self, token):
+        """Converts a token (str) in an id using the vocab."""
+        return self.sp_model.piece_to_id(token)
+
+    # Copied from transformers.models.llama.tokenization_llama.LlamaTokenizer._convert_id_to_token
+    def _convert_id_to_token(self, index):
+        """Converts an index (integer) in a token (str) using the vocab."""
+        token = self.sp_model.IdToPiece(index)
+        return token
+
+    def _decode(
+        self,
+        token_ids: List[int],
+        skip_special_tokens: bool = False,
+        spaces_between_special_tokens: bool = False,
+        **kwargs,
+    ) -> str:
+        sub_texts = []
+        current_sub_text = []
+        for ids in token_ids:
+            if skip_special_tokens and ids in self.all_special_ids:
+                continue
+            if ids in self._added_tokens_decoder:
+                if current_sub_text:
+                    sub_texts.append(self.sp_model.decode(current_sub_text))
+                sub_texts.append(self._added_tokens_decoder[ids].content)
+                current_sub_text = []
+            else:
+                current_sub_text.append(ids)
+        if current_sub_text:
+            sub_texts.append(self.sp_model.decode(current_sub_text))
+
+        if spaces_between_special_tokens:
+            sub_texts = " ".join(sub_texts)
+        else:
+            sub_texts = "".join(sub_texts)
+
+        return sub_texts
+
+    def convert_tokens_to_string(self, tokens):
+        """Converts a sequence of tokens (string) in a single string."""
+        current_sub_tokens = []
+        out_string = ""
+        for token in tokens:
+            # make sure that special tokens are not decoded using sentencepiece model
+            if token in self._added_tokens_encoder:
+                out_string += self.sp_model.decode(current_sub_tokens) + token
+                current_sub_tokens = []
+            else:
+                current_sub_tokens.append(token)
+        out_string += self.sp_model.decode(current_sub_tokens)
+        return out_string
+
+    # Copied from transformers.models.llama.tokenization_llama.LlamaTokenizer.save_vocabulary
+    def save_vocabulary(self, save_directory, filename_prefix: Optional[str] = None) -> Tuple[str]:
+        """
+        Save the vocabulary and special tokens file to a directory.
+
+        Args:
+            save_directory (`str`):
+                The directory in which to save the vocabulary.
+
+        Returns:
+            `Tuple(str)`: Paths to the files saved.
+        """
+        if not os.path.isdir(save_directory):
+            logger.error(f"Vocabulary path ({save_directory}) should be a directory")
+            return
+        out_vocab_file = os.path.join(
+            save_directory, (filename_prefix + "-" if filename_prefix else "") + VOCAB_FILES_NAMES["vocab_file"]
+        )
+
+        if os.path.abspath(self.vocab_file) != os.path.abspath(out_vocab_file) and os.path.isfile(self.vocab_file):
+            copyfile(self.vocab_file, out_vocab_file)
+        elif not os.path.isfile(self.vocab_file):
+            with open(out_vocab_file, "wb") as fi:
+                content_spiece_model = self.sp_model.serialized_model_proto()
+                fi.write(content_spiece_model)
+
+        return (out_vocab_file,)
+
+    # Copied from transformers.models.llama.tokenization_llama.LlamaTokenizer.build_inputs_with_special_tokens
+    def build_inputs_with_special_tokens(self, token_ids_0, token_ids_1=None):
+        bos_token_id = [self.bos_token_id] if self.add_bos_token else []
+        eos_token_id = [self.eos_token_id] if self.add_eos_token else []
+
+        output = bos_token_id + token_ids_0 + eos_token_id
+
+        if token_ids_1 is not None:
+            output = output + bos_token_id + token_ids_1 + eos_token_id
+
+        return output
+
+    # Copied from transformers.models.llama.tokenization_llama.LlamaTokenizer.get_special_tokens_mask
+    def get_special_tokens_mask(
+        self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None, already_has_special_tokens: bool = False
+    ) -> List[int]:
+        """
+        Retrieve sequence ids from a token list that has no special tokens added. This method is called when adding
+        special tokens using the tokenizer `prepare_for_model` method.
+
+        Args:
+            token_ids_0 (`List[int]`):
+                List of IDs.
+            token_ids_1 (`List[int]`, *optional*):
+                Optional second list of IDs for sequence pairs.
+            already_has_special_tokens (`bool`, *optional*, defaults to `False`):
+                Whether or not the token list is already formatted with special tokens for the model.
+
+        Returns:
+            `List[int]`: A list of integers in the range [0, 1]: 1 for a special token, 0 for a sequence token.
+        """
+        if already_has_special_tokens:
+            return super().get_special_tokens_mask(
+                token_ids_0=token_ids_0, token_ids_1=token_ids_1, already_has_special_tokens=True
+            )
+
+        bos_token_id = [1] if self.add_bos_token else []
+        eos_token_id = [1] if self.add_eos_token else []
+
+        if token_ids_1 is None:
+            return bos_token_id + ([0] * len(token_ids_0)) + eos_token_id
+        return (
+            bos_token_id
+            + ([0] * len(token_ids_0))
+            + eos_token_id
+            + bos_token_id
+            + ([0] * len(token_ids_1))
+            + eos_token_id
+        )
+
+    # Copied from transformers.models.llama.tokenization_llama.LlamaTokenizer.create_token_type_ids_from_sequences
+    def create_token_type_ids_from_sequences(
+        self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None
+    ) -> List[int]:
+        """
+        Creates a mask from the two sequences passed to be used in a sequence-pair classification task. An ALBERT
+        sequence pair mask has the following format:
+
+        ```
+        0 0 0 0 0 0 0 0 0 0 0 1 1 1 1 1 1 1 1 1
+        | first sequence    | second sequence |
+        ```
+
+        if token_ids_1 is None, only returns the first portion of the mask (0s).
+
+        Args:
+            token_ids_0 (`List[int]`):
+                List of ids.
+            token_ids_1 (`List[int]`, *optional*):
+                Optional second list of IDs for sequence pairs.
+
+        Returns:
+            `List[int]`: List of [token type IDs](../glossary#token-type-ids) according to the given sequence(s).
+        """
+        bos_token_id = [self.bos_token_id] if self.add_bos_token else []
+        eos_token_id = [self.eos_token_id] if self.add_eos_token else []
+
+        output = [0] * len(bos_token_id + token_ids_0 + eos_token_id)
+
+        if token_ids_1 is not None:
+            output += [1] * len(bos_token_id + token_ids_1 + eos_token_id)
+
+        return output

src/transformers/models/gemma/tokenization_gemma_fast.py

+# coding=utf-8
+# Copyright 2024 The HuggingFace Inc. team.
+#
+# 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.
+import os
+from shutil import copyfile
+from typing import Optional, Tuple
+
+from tokenizers import processors
+
+from ...tokenization_utils_fast import PreTrainedTokenizerFast
+from ...utils import is_sentencepiece_available, logging
+from ...utils.versions import require_version
+
+
+require_version("tokenizers>=0.13.3")
+
+if is_sentencepiece_available():
+    from .tokenization_gemma import GemmaTokenizer
+else:
+    GemmaTokenizer = None
+
+logger = logging.get_logger(__name__)
+VOCAB_FILES_NAMES = {"vocab_file": "tokenizer.model", "tokenizer_file": "tokenizer.json"}
+
+
+class GemmaTokenizerFast(PreTrainedTokenizerFast):
+    """
+    Construct a Gemma tokenizer fast. Based on byte-level Byte-Pair-Encoding.
+
+    This uses notably ByteFallback and no prefix space. Normalization is applied to replace  `" "` with `"▁"`
+
+    ```python
+    >>> from transformers import GemmaTokenizerFast
+
+    >>> tokenizer = GemmaTokenizerFast.from_pretrained("hf-internal-testing/dummy-gemma")
+    >>> tokenizer.encode("Hello this is a test")
+    [2, 4521, 736, 603, 476, 2121]
+    ```
+
+    If you want to change the `bos_token` or the `eos_token`, make sure to specify them when initializing the model, or
+    call `tokenizer.update_post_processor()` to make sure that the post-processing is correctly done (otherwise the
+    values of the first token and final token of an encoded sequence will not be correct). For more details, checkout
+    [post-processors] (https://huggingface.co/docs/tokenizers/api/post-processors) documentation.
+
+
+    This tokenizer inherits from [`PreTrainedTokenizerFast`] which contains most of the main methods. Users should
+    refer to this superclass for more information regarding those methods.
+
+    Args:
+        vocab_file (`str`, *optional*):
+            [SentencePiece](https://github.com/google/sentencepiece) file (generally has a .model extension) that
+            contains the vocabulary necessary to instantiate a tokenizer.
+        tokenizer_file (`str`, *optional*):
+            [tokenizers](https://github.com/huggingface/tokenizers) file (generally has a .json extension) that
+            contains everything needed to load the tokenizer.
+        clean_up_tokenization_spaces (`bool`, *optional*, defaults to `False`):
+            Whether or not to cleanup spaces after decoding, cleanup consists in removing potential artifacts like
+            extra spaces.
+        unk_token (`str` or `tokenizers.AddedToken`, *optional*, defaults to `"<unk>"`):
+            The unknown token. A token that is not in the vocabulary cannot be converted to an ID and is set to be this
+            token instead.
+        bos_token (`str` or `tokenizers.AddedToken`, *optional*, defaults to `"<bos>"`):
+            The beginning of sequence token that was used during pretraining. Can be used a sequence classifier token.
+        eos_token (`str` or `tokenizers.AddedToken`, *optional*, defaults to `"<eos>"`):
+            The end of sequence token.
+        pad_token (`str`, *optional*, defaults to `"<pad>"`):
+            The padding token
+        add_bos_token (`bool`, *optional*, defaults to `True`):
+            Whether or not to add an `bos_token` at the start of sequences.
+        add_eos_token (`bool`, *optional*, defaults to `False`):
+            Whether or not to add an `eos_token` at the end of sequences.
+    """
+
+    vocab_files_names = VOCAB_FILES_NAMES
+    slow_tokenizer_class = GemmaTokenizer
+    padding_side = "left"
+    model_input_names = ["input_ids", "attention_mask"]
+
+    def __init__(
+        self,
+        vocab_file=None,
+        tokenizer_file=None,
+        clean_up_tokenization_spaces=False,
+        unk_token="<unk>",
+        bos_token="<bos>",
+        eos_token="<eos>",
+        pad_token="<pad>",
+        add_bos_token=True,
+        add_eos_token=False,
+        **kwargs,
+    ):
+        super().__init__(
+            vocab_file=vocab_file,
+            tokenizer_file=tokenizer_file,
+            clean_up_tokenization_spaces=clean_up_tokenization_spaces,
+            unk_token=unk_token,
+            bos_token=bos_token,
+            eos_token=eos_token,
+            pad_token=pad_token,
+            add_bos_token=add_bos_token,
+            add_eos_token=add_eos_token,
+            **kwargs,
+        )
+        self._add_bos_token = add_bos_token
+        self._add_eos_token = add_eos_token
+        self.update_post_processor()
+        self.vocab_file = vocab_file
+
+    @property
+    def can_save_slow_tokenizer(self) -> bool:
+        return os.path.isfile(self.vocab_file) if self.vocab_file else False
+
+    # Copied from transformers.models.llama.tokenization_llama_fast.LlamaTokenizerFast.update_post_processor
+    def update_post_processor(self):
+        """
+        Updates the underlying post processor with the current `bos_token` and `eos_token`.
+        """
+        bos = self.bos_token
+        bos_token_id = self.bos_token_id
+        if bos is None and self.add_bos_token:
+            raise ValueError("add_bos_token = True but bos_token = None")
+
+        eos = self.eos_token
+        eos_token_id = self.eos_token_id
+        if eos is None and self.add_eos_token:
+            raise ValueError("add_eos_token = True but eos_token = None")
+
+        single = f"{(bos+':0 ') if self.add_bos_token else ''}$A:0{(' '+eos+':0') if self.add_eos_token else ''}"
+        pair = f"{single}{(' '+bos+':1') if self.add_bos_token else ''} $B:1{(' '+eos+':1') if self.add_eos_token else ''}"
+
+        special_tokens = []
+        if self.add_bos_token:
+            special_tokens.append((bos, bos_token_id))
+        if self.add_eos_token:
+            special_tokens.append((eos, eos_token_id))
+        self._tokenizer.post_processor = processors.TemplateProcessing(
+            single=single, pair=pair, special_tokens=special_tokens
+        )
+
+    @property
+    def add_eos_token(self):
+        return self._add_eos_token
+
+    @property
+    def add_bos_token(self):
+        return self._add_bos_token
+
+    @add_eos_token.setter
+    def add_eos_token(self, value):
+        self._add_eos_token = value
+        self.update_post_processor()
+
+    @add_bos_token.setter
+    def add_bos_token(self, value):
+        self._add_bos_token = value
+        self.update_post_processor()
+
+    # Copied from transformers.models.llama.tokenization_llama_fast.LlamaTokenizerFast.save_vocabulary
+    def save_vocabulary(self, save_directory: str, filename_prefix: Optional[str] = None) -> Tuple[str]:
+        if not self.can_save_slow_tokenizer:
+            raise ValueError(
+                "Your fast tokenizer does not have the necessary information to save the vocabulary for a slow "
+                "tokenizer."
+            )
+
+        if not os.path.isdir(save_directory):
+            logger.error(f"Vocabulary path ({save_directory}) should be a directory")
+            return
+        out_vocab_file = os.path.join(
+            save_directory, (filename_prefix + "-" if filename_prefix else "") + VOCAB_FILES_NAMES["vocab_file"]
+        )
+
+        if os.path.abspath(self.vocab_file) != os.path.abspath(out_vocab_file):
+            copyfile(self.vocab_file, out_vocab_file)
+
+        return (out_vocab_file,)
+
+    # Copied from transformers.models.llama.tokenization_llama_fast.LlamaTokenizerFast.build_inputs_with_special_tokens
+    def build_inputs_with_special_tokens(self, token_ids_0, token_ids_1=None):
+        bos_token_id = [self.bos_token_id] if self.add_bos_token else []
+        eos_token_id = [self.eos_token_id] if self.add_eos_token else []
+
+        output = bos_token_id + token_ids_0 + eos_token_id
+
+        if token_ids_1 is not None:
+            output = output + bos_token_id + token_ids_1 + eos_token_id
+
+        return output

src/transformers/utils/dummy_flax_objects.py

@@ -737,6 +737,27 @@ class FlaxEncoderDecoderModel(metaclass=DummyObject):
         requires_backends(self, ["flax"])
 
 
+class FlaxGemmaForCausalLM(metaclass=DummyObject):
+    _backends = ["flax"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["flax"])
+
+
+class FlaxGemmaModel(metaclass=DummyObject):
+    _backends = ["flax"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["flax"])
+
+
+class FlaxGemmaPreTrainedModel(metaclass=DummyObject):
+    _backends = ["flax"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["flax"])
+
+
 class FlaxGPT2LMHeadModel(metaclass=DummyObject):
     _backends = ["flax"]
 

src/transformers/utils/dummy_pt_objects.py

@@ -3811,6 +3811,34 @@ class FuyuPreTrainedModel(metaclass=DummyObject):
         requires_backends(self, ["torch"])
 
 
+class GemmaForCausalLM(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
+class GemmaForSequenceClassification(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
+class GemmaModel(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
+class GemmaPreTrainedModel(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
 GIT_PRETRAINED_MODEL_ARCHIVE_LIST = None
 
 

src/transformers/utils/dummy_sentencepiece_objects.py

@@ -79,6 +79,13 @@ class FNetTokenizer(metaclass=DummyObject):
         requires_backends(self, ["sentencepiece"])
 
 
+class GemmaTokenizer(metaclass=DummyObject):
+    _backends = ["sentencepiece"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["sentencepiece"])
+
+
 class GPTSw3Tokenizer(metaclass=DummyObject):
     _backends = ["sentencepiece"]
 

src/transformers/utils/dummy_tokenizers_objects.py

@@ -170,6 +170,13 @@ class FunnelTokenizerFast(metaclass=DummyObject):
         requires_backends(self, ["tokenizers"])
 
 
+class GemmaTokenizerFast(metaclass=DummyObject):
+    _backends = ["tokenizers"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["tokenizers"])
+
+
 class GPT2TokenizerFast(metaclass=DummyObject):
     _backends = ["tokenizers"]
 

tests/models/gemma/test_modeling_flax_gemma.py

+# Copyright 2024 The HuggingFace 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.
+
+
+import unittest
+
+import numpy as np
+
+from transformers import AutoTokenizer, GemmaConfig, is_flax_available
+from transformers.testing_utils import require_flax, slow
+
+from ...generation.test_flax_utils import FlaxGenerationTesterMixin
+from ...test_modeling_flax_common import FlaxModelTesterMixin, ids_tensor
+
+
+if is_flax_available():
+    import jax
+    import jax.numpy as jnp
+
+    from transformers.models.gemma.modeling_flax_gemma import (
+        FlaxGemmaForCausalLM,
+        FlaxGemmaModel,
+    )
+
+
+class FlaxGemmaModelTester:
+    def __init__(
+        self,
+        parent,
+        batch_size=2,
+        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,
+        num_key_value_heads=2,
+        intermediate_size=37,
+        hidden_act="gelu",
+        hidden_dropout_prob=0.1,
+        attention_probs_dropout_prob=0.1,
+        max_position_embeddings=512,
+        initializer_range=0.02,
+    ):
+        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.initializer_range = initializer_range
+        self.scope = None
+        self.bos_token_id = vocab_size - 1
+        self.eos_token_id = vocab_size - 1
+        self.pad_token_id = 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 = np.tril(np.ones((self.batch_size, self.seq_length)))
+
+        config = GemmaConfig(
+            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,
+            head_dim=self.hidden_size // 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,
+            use_cache=True,
+            is_decoder=False,
+            initializer_range=self.initializer_range,
+        )
+
+        return config, input_ids, input_mask
+
+    def prepare_config_and_inputs_for_common(self):
+        config_and_inputs = self.prepare_config_and_inputs()
+        config, input_ids, attention_mask = config_and_inputs
+        inputs_dict = {"input_ids": input_ids, "attention_mask": attention_mask}
+        return config, inputs_dict
+
+    def check_use_cache_forward(self, model_class_name, config, input_ids, attention_mask):
+        max_decoder_length = 20
+        model = model_class_name(config)
+
+        past_key_values = model.init_cache(input_ids.shape[0], max_decoder_length)
+        attention_mask = jnp.ones((input_ids.shape[0], max_decoder_length), dtype="i4")
+
+        position_ids = jnp.broadcast_to(
+            jnp.arange(input_ids.shape[-1] - 1)[None, :], (input_ids.shape[0], input_ids.shape[-1] - 1)
+        )
+        outputs_cache = model(
+            input_ids[:, :-1],
+            attention_mask=attention_mask,
+            past_key_values=past_key_values,
+            position_ids=position_ids,
+        )
+
+        position_ids = jnp.array(input_ids.shape[0] * [[input_ids.shape[-1] - 1]], dtype="i4")
+        outputs_cache_next = model(
+            input_ids[:, -1:],
+            attention_mask=attention_mask,
+            past_key_values=outputs_cache.past_key_values,
+            position_ids=position_ids,
+        )
+
+        outputs = model(input_ids)
+
+        diff = np.max(np.abs((outputs_cache_next[0][:, -1, :5] - outputs[0][:, -1, :5])))
+        self.parent.assertTrue(diff < 1e-3, msg=f"Max diff is {diff}")
+
+    def check_use_cache_forward_with_attn_mask(self, model_class_name, config, input_ids, attention_mask):
+        max_decoder_length = 20
+        model = model_class_name(config)
+
+        attention_mask_cache = jnp.concatenate(
+            [attention_mask, jnp.zeros((attention_mask.shape[0], max_decoder_length - attention_mask.shape[1]))],
+            axis=-1,
+        )
+
+        past_key_values = model.init_cache(input_ids.shape[0], max_decoder_length)
+        position_ids = jnp.broadcast_to(
+            jnp.arange(input_ids.shape[-1] - 1)[None, :], (input_ids.shape[0], input_ids.shape[-1] - 1)
+        )
+
+        outputs_cache = model(
+            input_ids[:, :-1],
+            attention_mask=attention_mask_cache,
+            past_key_values=past_key_values,
+            position_ids=position_ids,
+        )
+        position_ids = jnp.array(input_ids.shape[0] * [[input_ids.shape[-1] - 1]], dtype="i4")
+        outputs_cache_next = model(
+            input_ids[:, -1:],
+            past_key_values=outputs_cache.past_key_values,
+            attention_mask=attention_mask_cache,
+            position_ids=position_ids,
+        )
+
+        outputs = model(input_ids, attention_mask=attention_mask)
+
+        diff = np.max(np.abs((outputs_cache_next[0][:, -1, :5] - outputs[0][:, -1, :5])))
+        self.parent.assertTrue(diff < 1e-3, msg=f"Max diff is {diff}")
+
+
+@require_flax
+class FlaxGemmaModelTest(FlaxModelTesterMixin, FlaxGenerationTesterMixin, unittest.TestCase):
+    all_model_classes = (FlaxGemmaModel, FlaxGemmaForCausalLM) if is_flax_available() else ()
+    all_generative_model_classes = (FlaxGemmaForCausalLM,) if is_flax_available() else ()
+
+    def setUp(self):
+        self.model_tester = FlaxGemmaModelTester(self)
+
+    def test_use_cache_forward(self):
+        for model_class_name in self.all_model_classes:
+            config, input_ids, attention_mask = self.model_tester.prepare_config_and_inputs()
+            self.model_tester.check_use_cache_forward(model_class_name, config, input_ids, attention_mask)
+
+    def test_use_cache_forward_with_attn_mask(self):
+        for model_class_name in self.all_model_classes:
+            config, input_ids, attention_mask = self.model_tester.prepare_config_and_inputs()
+            self.model_tester.check_use_cache_forward_with_attn_mask(
+                model_class_name, config, input_ids, attention_mask
+            )
+
+    @slow
+    def test_model_from_pretrained(self):
+        for model_class_name in self.all_model_classes:
+            model = model_class_name.from_pretrained("google/gemma-2b", from_pt=True)
+            outputs = model(np.ones((1, 1)))
+            self.assertIsNotNone(outputs)
+
+
+@slow
+@require_flax
+class FlaxGemmaIntegrationTest(unittest.TestCase):
+    input_text = ["The capital of France is", "To play the perfect cover drive"]
+    model_id = "google/gemma-2b"
+    revision = "flax"
+
+    def setUp(self):
+        self.model, self.params = FlaxGemmaForCausalLM.from_pretrained(
+            self.model_id, revision=self.revision, _do_init=False
+        )
+        self.tokenizer = AutoTokenizer.from_pretrained(self.model_id)
+        self.tokenizer.padding_side = "left"
+
+    def test_logits(self):
+        inputs = self.tokenizer(self.input_text, return_tensors="np", padding=True)
+        # fmt: off
+        EXPECTED_MEAN = [
+            [-16.427, -21.386, -35.491, -36.258, -31.401, -36.370, -37.598],
+            [-21.386, -32.150, -33.155, -34.344, -34.706, -34.678, -38.495],
+        ]
+        EXPECTED_SLICE = [-33.462, -16.481, -30.837, -32.195, -33.113]
+        # fmt: on
+
+        logits = self.model(**inputs, params=self.params).logits
+
+        diff_mean = jnp.abs(logits.mean(-1) - np.array(EXPECTED_MEAN)).max()
+        diff_slice = jnp.abs(logits[0, -1, 475:480] - np.array(EXPECTED_SLICE)).max()
+
+        self.assertAlmostEqual(diff_mean, 0, places=3)
+        self.assertAlmostEqual(diff_slice, 0, places=3)
+
+    def test_generation(self):
+        EXPECTED_TEXTS = [
+            "The capital of France is a city of contrasts. It is a city of history, of art, of culture, of fashion",
+            "To play the perfect cover drive, you need to have a good technique and a good mindset.\n\nThe cover drive is a shot",
+        ]
+        inputs = self.tokenizer(self.input_text, return_tensors="np", padding=True)
+
+        output = self.model.generate(**inputs, params=self.params, max_new_tokens=20, do_sample=False)
+        output_text = self.tokenizer.batch_decode(output.sequences, skip_special_tokens=True)
+
+        self.assertEqual(output_text, EXPECTED_TEXTS)
+
+    def test_jit_generation(self):
+        EXPECTED_TEXTS = [
+            "The capital of France is a city of contrasts. It is a city of history, culture, and art, but it is",
+            "To play the perfect cover drive, you need to have a good technique and a good mindset.\n\nThe cover drive is a shot",
+        ]
+        inputs = self.tokenizer(self.input_text, return_tensors="np", padding=True)
+
+        def generate(input_ids, attention_mask):
+            outputs = self.model.generate(
+                input_ids, attention_mask=attention_mask, params=self.params, max_new_tokens=20, do_sample=False
+            )
+            return outputs
+
+        jit_generate = jax.jit(generate)
+        output_sequences = jit_generate(**inputs).sequences
+        output_text = self.tokenizer.batch_decode(output_sequences, skip_special_tokens=True)
+
+        self.assertEqual(output_text, EXPECTED_TEXTS)

utils/check_docstrings.py

@@ -233,6 +233,8 @@ OBJECTS_TO_IGNORE = [
     "FlaxGPTNeoModel",
     "FlaxLlamaForCausalLM",
     "FlaxLlamaModel",
+    "FlaxGemmaForCausalLM",
+    "FlaxGemmaModel",
     "FlaxMBartForConditionalGeneration",
     "FlaxMBartForQuestionAnswering",
     "FlaxMBartForSequenceClassification",

utils/not_doctested.txt

@@ -575,6 +575,10 @@ src/transformers/models/funnel/convert_funnel_original_tf_checkpoint_to_pytorch.
 src/transformers/models/funnel/modeling_funnel.py
 src/transformers/models/funnel/modeling_tf_funnel.py
 src/transformers/models/fuyu/convert_fuyu_model_weights_to_hf.py
+src/transformers/models/gemma/configuration_gemma.py
+src/transformers/models/gemma/convert_gemma_weights_to_hf.py
+src/transformers/models/gemma/modeling_flax_gemma.py
+src/transformers/models/gemma/modeling_gemma.py
 src/transformers/models/git/configuration_git.py
 src/transformers/models/git/convert_git_to_pytorch.py
 src/transformers/models/glpn/configuration_glpn.py