Encoder-Decoder Gemma (#38332)

* Initial submit

* Fix bugs:
1. add __init__ file
2. tied word embedding
3. support flash/flex attention
4. model saving and loading

* Code refactor:
* Rename encdecgemma to t5gemma.
* Split attention into self- and cross-attention
* Split stack into encoder and decoder
* Add test cases
* Add auto configuration

* Update configurations.

* Fix bugs related to copy and attribute checks

* Fix type union

* Fix merge errors

* run ruff format

* Run make style and update tests.

* Add t5gemma model doc.

* ruff and style formatting.

* Add missed module config.

* Add dummy checkpoint link to pass tests (need updated when real checkpoints are uplioaded.).

* Update model doc.

* Minor updates following Arthur's comments:
* replace docstrings with auto_docstrings
* remove checkpoint layers
* remove deprecate_kwargs

* fix rebase errors

* Fix docstring issues.

* fix t5gemma doc issue.

* run ruff format

* Updates:
* split encoder-only model out
* make t5gemmamodel encoder-decoder only
* update token and sequence classification
* update tests

docs/source/en/_toctree.yml

@@ -655,6 +655,8 @@
         title: SwitchTransformers
       - local: model_doc/t5
         title: T5
+      - local: model_doc/t5gemma
+        title: T5Gemma
       - local: model_doc/t5v1.1
         title: T5v1.1
       - local: model_doc/tapex

docs/source/en/model_doc/t5gemma.md

@@ -0,0 +1,107 @@
+
+<!--Copyright 2025 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.
+
+⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+
+-->
+
+
+# T5Gemma
+
+T5Gemma (aka encoder-decoder Gemma) was proposed in a [research paper](https://arxiv.org/abs/2504.06225) by Google. It is a family of encoder-decoder large langauge models, developed by adapting pretrained decoder-only models into encoder-decoder. T5Gemma includes pretrained and instruction-tuned variants. The architecture is based on transformer encoder-decoder design following T5, with improvements from Gemma 2: GQA, RoPE, GeGLU activation, RMSNorm, and interleaved local/global attention.
+
+T5Gemma has two groups of model sizes: 1) [Gemma 2](https://ai.google.dev/gemma/docs/core/model_card_2) sizes (2B-2B, 9B-2B, and 9B-9B), which are based on the offical Gemma 2 models (2B and 9B); and 2) [T5](https://arxiv.org/abs/1910.10683) sizes (Small, Base, Large, and XL), where are pretrained under the Gemma 2 framework following T5 configuration. In addition, we also provide a model at ML size (medium large, ~2B in total), which is in-between T5 Large and T5 XL.
+
+The pretrained varaints are trained with two objectives: prefix language modeling with knowledge distillation (PrefixLM) and UL2, separately. We release both variants for each model size. The instruction-turned varaints was post-trained with supervised fine-tuning and reinforcement learning.
+
+The example below demonstrates how to chat with the model with [`Pipeline`] or the [`AutoModel`] class, and from the command line.
+
+<hfoptions id="usage">
+<hfoption id="Pipeline">
+
+
+```python
+import torch
+from transformers import pipeline
+
+pipe = pipeline(
+    task="text2text-generation",
+    model="google/t5gemma-placeholder",
+    torch_dtype=torch.bfloat16,
+    device="cuda",
+)
+
+pipe("Question: Why is the sky blue?\nAnswer:", max_new_tokens=50)
+```
+
+</hfoption>
+<hfoption id="AutoModel">
+
+```python
+import torch
+from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
+
+tokenizer = AutoTokenizer.from_pretrained("google/t5gemma-placeholder")
+model = AutoModelForSeq2SeqLM.from_pretrained(
+    "google/t5gemma-placeholder",
+    torch_dtype=torch.bfloat16,
+    device_map="auto"
+)
+
+input_text = "Question: Why is the sky blue?\nAnswer:"
+input_ids = tokenizer(input_text, return_tensors="pt").to("cuda")
+
+outputs = model.generate(**input_ids, max_new_tokens=32)
+print(tokenizer.decode(outputs[0], skip_special_tokens=True))
+
+```
+
+</hfoption>
+<hfoption id="transformers CLI">
+
+```
+echo -e "Question: Why is the sky blue? Answer:" | transformers run --task text2text-generation --model google/t5gemma-placeholder --device 0
+```
+
+## T5GemmaConfig
+
+[[autodoc]] T5GemmaConfig
+
+## T5GemmaModuleConfig
+
+[[autodoc]] T5GemmaModuleConfig
+
+## T5GemmaModel
+
+[[autodoc]] T5GemmaModel
+    - forward
+
+## T5GemmaEncoderModel
+
+[[autodoc]] T5GemmaEncoderModel
+    - forward
+
+## T5GemmaForConditionalGeneration
+
+[[autodoc]] T5GemmaForConditionalGeneration
+    - forward
+
+## T5GemmaForSequenceClassification
+
+[[autodoc]] T5GemmaForSequenceClassification
+    - forward
+
+## T5GemmaForTokenClassification
+
+[[autodoc]] T5GemmaForTokenClassification
+    - forward

src/transformers/models/__init__.py

@@ -294,6 +294,7 @@ if TYPE_CHECKING:
     from .swinv2 import *
     from .switch_transformers import *
     from .t5 import *
+    from .t5gemma import *
     from .table_transformer import *
     from .tapas import *
     from .textnet import *

src/transformers/models/auto/configuration_auto.py

@@ -333,6 +333,7 @@ CONFIG_MAPPING_NAMES = OrderedDict[str, str](
         ("swinv2", "Swinv2Config"),
         ("switch_transformers", "SwitchTransformersConfig"),
         ("t5", "T5Config"),
+        ("t5gemma", "T5GemmaConfig"),
         ("table-transformer", "TableTransformerConfig"),
         ("tapas", "TapasConfig"),
         ("textnet", "TextNetConfig"),
@@ -721,6 +722,7 @@ MODEL_NAMES_MAPPING = OrderedDict[str, str](
         ("swinv2", "Swin Transformer V2"),
         ("switch_transformers", "SwitchTransformers"),
         ("t5", "T5"),
+        ("t5gemma", "T5Gemma"),
         ("t5v1.1", "T5v1.1"),
         ("table-transformer", "Table Transformer"),
         ("tapas", "TAPAS"),

src/transformers/models/auto/modeling_auto.py

@@ -310,6 +310,7 @@ MODEL_MAPPING_NAMES = OrderedDict(
         ("swinv2", "Swinv2Model"),
         ("switch_transformers", "SwitchTransformersModel"),
         ("t5", "T5Model"),
+        ("t5gemma", "T5GemmaModel"),
         ("table-transformer", "TableTransformerModel"),
         ("tapas", "TapasModel"),
         ("textnet", "TextNetModel"),
@@ -430,6 +431,7 @@ MODEL_FOR_PRETRAINING_MAPPING_NAMES = OrderedDict(
         ("squeezebert", "SqueezeBertForMaskedLM"),
         ("switch_transformers", "SwitchTransformersForConditionalGeneration"),
         ("t5", "T5ForConditionalGeneration"),
+        ("t5gemma", "T5GemmaForConditionalGeneration"),
         ("tapas", "TapasForMaskedLM"),
         ("transfo-xl", "TransfoXLLMHeadModel"),
         ("tvlt", "TvltForPreTraining"),
@@ -524,6 +526,7 @@ MODEL_WITH_LM_HEAD_MAPPING_NAMES = OrderedDict(
         ("squeezebert", "SqueezeBertForMaskedLM"),
         ("switch_transformers", "SwitchTransformersForConditionalGeneration"),
         ("t5", "T5ForConditionalGeneration"),
+        ("t5gemma", "T5GemmaForConditionalGeneration"),
         ("tapas", "TapasForMaskedLM"),
         ("transfo-xl", "TransfoXLLMHeadModel"),
         ("wav2vec2", "Wav2Vec2ForMaskedLM"),
@@ -1044,6 +1047,7 @@ MODEL_FOR_SEQ_TO_SEQ_CAUSAL_LM_MAPPING_NAMES = OrderedDict(
         ("seamless_m4t_v2", "SeamlessM4Tv2ForTextToText"),
         ("switch_transformers", "SwitchTransformersForConditionalGeneration"),
         ("t5", "T5ForConditionalGeneration"),
+        ("t5gemma", "T5GemmaForConditionalGeneration"),
         ("umt5", "UMT5ForConditionalGeneration"),
         ("xlm-prophetnet", "XLMProphetNetForConditionalGeneration"),
     ]
@@ -1156,6 +1160,7 @@ MODEL_FOR_SEQUENCE_CLASSIFICATION_MAPPING_NAMES = OrderedDict(
         ("stablelm", "StableLmForSequenceClassification"),
         ("starcoder2", "Starcoder2ForSequenceClassification"),
         ("t5", "T5ForSequenceClassification"),
+        ("t5gemma", "T5GemmaForSequenceClassification"),
         ("tapas", "TapasForSequenceClassification"),
         ("transfo-xl", "TransfoXLForSequenceClassification"),
         ("umt5", "UMT5ForSequenceClassification"),
@@ -1349,6 +1354,7 @@ MODEL_FOR_TOKEN_CLASSIFICATION_MAPPING_NAMES = OrderedDict(
         ("stablelm", "StableLmForTokenClassification"),
         ("starcoder2", "Starcoder2ForTokenClassification"),
         ("t5", "T5ForTokenClassification"),
+        ("t5gemma", "T5GemmaForTokenClassification"),
         ("umt5", "UMT5ForTokenClassification"),
         ("xlm", "XLMForTokenClassification"),
         ("xlm-roberta", "XLMRobertaForTokenClassification"),
@@ -1582,6 +1588,7 @@ MODEL_FOR_TEXT_ENCODING_MAPPING_NAMES = OrderedDict(
         ("roformer", "RoFormerModel"),
         ("squeezebert", "SqueezeBertModel"),
         ("t5", "T5EncoderModel"),
+        ("t5gemma", "T5GemmaEncoderModel"),
         ("umt5", "UMT5EncoderModel"),
         ("xlm", "XLMModel"),
         ("xlm-roberta", "XLMRobertaModel"),

src/transformers/models/auto/tokenization_auto.py

@@ -582,6 +582,13 @@ TOKENIZER_MAPPING_NAMES = OrderedDict[str, tuple[Optional[str], Optional[str]]](
                 "T5TokenizerFast" if is_tokenizers_available() else None,
             ),
         ),
+        (
+            "t5gemma",
+            (
+                "GemmaTokenizer" if is_sentencepiece_available() else None,
+                "GemmaTokenizerFast" if is_tokenizers_available() else None,
+            ),
+        ),
         ("tapas", ("TapasTokenizer", None)),
         ("tapex", ("TapexTokenizer", None)),
         ("transfo-xl", ("TransfoXLTokenizer", None)),

src/transformers/models/t5gemma/__init__.py

@@ -0,0 +1,27 @@
+# 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.
+from typing import TYPE_CHECKING
+
+from ...utils import _LazyModule
+from ...utils.import_utils import define_import_structure
+
+
+if TYPE_CHECKING:
+    from .configuration_encdecgemma2 import *
+    from .modeling_encdecgemma2 import *
+else:
+    import sys
+
+    _file = globals()["__file__"]
+    sys.modules[__name__] = _LazyModule(__name__, _file, define_import_structure(_file), module_spec=__spec__)

src/transformers/models/t5gemma/configuration_t5gemma.py

@@ -0,0 +1,333 @@
+#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
+#           This file was automatically generated from src/transformers/models/t5gemma/modular_t5gemma.py.
+#               Do NOT edit this file manually as any edits will be overwritten by the generation of
+#             the file from the modular. If any change should be done, please apply the change to the
+#                          modular_t5gemma.py file directly. One of our CI enforces this.
+#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
+# coding=utf-8
+# Copyright 2025 Google Inc. 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 Any, Optional, Union
+
+from ...configuration_utils import PretrainedConfig, layer_type_validation
+
+
+class T5GemmaModuleConfig(PretrainedConfig):
+    r"""
+        This is the configuration class to store the configuration of a [`T5GemmaModuleModel`]. It is used to instantiate an T5GemmaModule
+        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 T5GemmaModule-7B.
+        e.g. [google/t5_gemma_module-7b](https://huggingface.co/google/t5_gemma_module-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 T5GemmaModule model. Defines the number of different tokens that can be represented by the
+                `inputs_ids` passed when calling [`T5GemmaModuleModel`]
+            hidden_size (`int`, *optional*, defaults to 2304):
+                Dimension of the hidden representations.
+            intermediate_size (`int`, *optional*, defaults to 9216):
+                Dimension of the MLP representations.
+            num_hidden_layers (`int`, *optional*, defaults to 26):
+                Number of hidden layers in the Transformer decoder.
+            num_attention_heads (`int`, *optional*, defaults to 8):
+                Number of attention heads for each attention layer in the Transformer decoder.
+            num_key_value_heads (`int`, *optional*, defaults to 4):
+                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, check out [this
+                paper](https://huggingface.co/papers/2305.13245). If it is not specified, will default to
+                `num_attention_heads`.
+            head_dim (`int`, *optional*, defaults to 256):
+                The attention head dimension.
+            hidden_activation (`str` or `function`, *optional*, defaults to `"gelu_pytorch_tanh"`):
+                The non-linear activation function (function or string) in the decoder. Will default to `"gelu_pytorch_tanh"`
+                if not specified. `"gelu_pytorch_tanh"` uses an approximation of the `"gelu"` activation function.
+            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.
+            query_pre_attn_scalar (`float`, *optional*, defaults to 256):
+                scaling factor used on the attention scores
+            sliding_window (`int`, *optional*, defaults to 4096):
+                in T5GemmaModule, every other layer uses sliding window attention. This is the size of the sliding window.
+            layer_types (`list`, *optional*):
+                Attention pattern for each layer.
+            final_logit_softcapping (`float`, *optional*, defaults to 30.0):
+                scaling factor when applying tanh softcapping on the logits.
+            attn_logit_softcapping (`float`, *optional*, defaults to 50.0):
+                scaling factor when applying tanh softcapping on the attention scores.
+
+        ```python
+        >>> from transformers import T5GemmaModuleModel, T5GemmaModuleConfig
+        >>> # Initializing a T5GemmaModule t5_gemma_module-7b style configuration
+        >>> configuration = T5GemmaModuleConfig()
+        >>> # Initializing a model from the t5_gemma_module-7b style configuration
+        >>> model = T5GemmaModuleModel(configuration)
+        >>> # Accessing the model configuration
+        >>> configuration = model.config
+        ```
+    Module config (encoder or decoder): the same as Gemma2Config."""
+
+    model_type = "t5_gemma_module"
+    keys_to_ignore_at_inference = ["past_key_values"]
+    base_model_tp_plan = {
+        "layers.*.self_attn.q_proj": "colwise",
+        "layers.*.self_attn.k_proj": "colwise",
+        "layers.*.self_attn.v_proj": "colwise",
+        "layers.*.self_attn.o_proj": "rowwise",
+        "layers.*.mlp.gate_proj": "colwise",
+        "layers.*.mlp.up_proj": "colwise",
+        "layers.*.mlp.down_proj": "rowwise",
+    }
+    base_model_pp_plan = {
+        "embed_tokens": (["input_ids"], ["inputs_embeds"]),
+        "layers": (["hidden_states", "attention_mask"], ["hidden_states"]),
+        "norm": (["hidden_states"], ["hidden_states"]),
+    }
+
+    def __init__(
+        self,
+        vocab_size=256000,
+        hidden_size=2304,
+        intermediate_size=9216,
+        num_hidden_layers=26,
+        num_attention_heads=8,
+        num_key_value_heads=4,
+        head_dim=256,
+        hidden_activation="gelu_pytorch_tanh",
+        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,
+        query_pre_attn_scalar=256,
+        sliding_window=4096,
+        layer_types=None,
+        final_logit_softcapping=30.0,
+        attn_logit_softcapping=50.0,
+        **kwargs,
+    ):
+        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,
+        )
+        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.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
+        self.hidden_activation = hidden_activation
+        self.query_pre_attn_scalar = query_pre_attn_scalar
+        self.sliding_window = sliding_window
+        self.final_logit_softcapping = final_logit_softcapping
+        self.attn_logit_softcapping = attn_logit_softcapping
+        self.layer_types = layer_types
+
+        if self.layer_types is None:
+            self.layer_types = [
+                "sliding_attention" if bool((i + 1) % 2) else "full_attention" for i in range(self.num_hidden_layers)
+            ]
+        layer_type_validation(self.layer_types)
+
+
+class T5GemmaConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`T5GemmaModel`]. It is used to instantiate an T5Gemma
+    model according to the specified arguments, defining the model architecture. Instantiating a configuration with the
+    defaults will yield a similar configuration to a hypothetical balanced Gemma2 encoder-decoder model.
+    e.g. [google/t5gemma-placeholder](https://huggingface.co/google/t5gemma-placeholder)
+    ```python
+    >>> from transformers import T5GemmaConfig, T5GemmaModel
+    >>> t5gemma_config = T5GemmaConfig.from_pretrained("google/t5gemma-placeholder")
+    >>> model = T5GemmaModel(t5gemma_config)
+    ```
+    Configuration objects inherit from [PretrainedConfig] and can be used to control the model outputs. Read the
+    documentation from [PretrainedConfig] for more information.
+    Args:
+        encoder (`Union[T5GemmaModuleConfig, dict]`, optional, *optional*):
+            Configuration for the encoder.
+        decoder (`Union[T5GemmaModuleConfig, dict]`, optional, *optional*):
+            Configuration for the decoder.
+        is_encoder_decoder (bool, optional, *optional*, defaults to `True`):
+            Whether the model is used as an encoder/decoder or not.
+        dropout_rate (`float`, *optional*, defaults to 0.0):
+            The ratio for all dropout layers (following T5).
+        classifier_dropout_rate (`float`, *optional*, defaults to 0.0):
+            The dropout ratio for classifier (following T5).
+        attention_dropout (`float`, *optional*, defaults to 0.0):
+            The dropout ratio for attention.
+        tie_word_embeddings (`bool`, *optional*, defaults to `True`):
+            Whether tie input and output embeddings.
+        kwargs (additional keyword arguments, optional, *optional*):
+            Will be passed to the PretrainedConfig base class.
+    """
+
+    model_type = "t5gemma"
+    keys_to_ignore_at_inference = ["past_key_values"]
+    base_model_tp_plan = {
+        # encoder
+        "encoder.layers.*.self_attn.q_proj": "colwise",
+        "encoder.layers.*.self_attn.k_proj": "colwise",
+        "encoder.layers.*.self_attn.v_proj": "colwise",
+        "encoder.layers.*.self_attn.o_proj": "rowwise",
+        "encoder.layers.*.mlp.gate_proj": "colwise",
+        "encoder.layers.*.mlp.up_proj": "colwise",
+        "encoder.layers.*.mlp.down_proj": "rowwise",
+        # decoder
+        "decoder.layers.*.self_attn.q_proj": "colwise",
+        "decoder.layers.*.self_attn.k_proj": "colwise",
+        "decoder.layers.*.self_attn.v_proj": "colwise",
+        "decoder.layers.*.self_attn.o_proj": "rowwise",
+        "decoder.layers.*.cross_attn.q_proj": "colwise",
+        "decoder.layers.*.cross_attn.k_proj": "colwise",
+        "decoder.layers.*.cross_attn.v_proj": "colwise",
+        "decoder.layers.*.cross_attn.o_proj": "rowwise",
+        "decoder.layers.*.mlp.gate_proj": "colwise",
+        "decoder.layers.*.mlp.up_proj": "colwise",
+        "decoder.layers.*.mlp.down_proj": "rowwise",
+    }
+    base_model_pp_plan = {
+        # encoder
+        "encoder.embed_tokens": (["input_ids"], ["inputs_embeds"]),
+        "encoder.layers": (["hidden_states", "attention_mask"], ["hidden_states"]),
+        "encoder.norm": (["hidden_states"], ["hidden_states"]),
+        # decoder
+        "decoder.embed_tokens": (["input_ids"], ["inputs_embeds"]),
+        "decoder.layers": (["hidden_states", "attention_mask"], ["hidden_states"]),
+        "decoder.norm": (["hidden_states"], ["hidden_states"]),
+    }
+
+    def __init__(
+        self,
+        encoder: Optional[Union[T5GemmaModuleConfig, dict[Any, Any]]] = None,
+        decoder: Optional[Union[T5GemmaModuleConfig, dict[Any, Any]]] = None,
+        is_encoder_decoder: bool = True,
+        dropout_rate: float = 0.0,
+        classifier_dropout_rate: float = 0.0,
+        attention_dropout: float = 0.0,
+        tie_word_embeddings: bool = True,
+        **kwargs,
+    ):
+        # Encoder.
+        if isinstance(encoder, dict):
+            # From preset configuration
+            encoder = T5GemmaModuleConfig(**encoder)
+        elif encoder is None:
+            # From scratch
+            encoder = T5GemmaModuleConfig()
+        else:
+            assert isinstance(encoder, T5GemmaModuleConfig), f"{type(encoder)} is not supported."
+
+        # Decoder.
+        if isinstance(decoder, dict):
+            # From preset configuration
+            decoder = T5GemmaModuleConfig(**decoder)
+        elif decoder is None:
+            # From scratch
+            decoder = encoder
+        else:
+            assert isinstance(decoder, T5GemmaModuleConfig), f"{type(decoder)} is not supported."
+
+        # Decouple encoder and decoder config in any case
+        encoder = T5GemmaModuleConfig(**encoder.to_dict())
+        decoder = T5GemmaModuleConfig(**decoder.to_dict())
+
+        encoder.is_decoder = False
+        encoder.dropout_rate = dropout_rate
+        encoder.attention_dropout = attention_dropout
+        self.encoder = encoder
+
+        decoder.is_decoder = True
+        decoder.use_cache = True
+        decoder.dropout_rate = dropout_rate
+        decoder.attention_dropout = attention_dropout
+        decoder.cross_attention_hidden_size = encoder.hidden_size
+        self.decoder = decoder
+
+        for special_token_key in ["bos_token_id", "pad_token_id", "eos_token_id"]:
+            if special_token_key not in kwargs:
+                kwargs[special_token_key] = getattr(decoder, special_token_key)
+
+        super().__init__(**kwargs)
+
+        self.is_encoder_decoder = is_encoder_decoder
+        self.use_cache = kwargs.get("use_cache", decoder.use_cache)
+        self.initializer_range = kwargs.get("initializer_range", decoder.initializer_range)
+        self.dropout_rate = dropout_rate
+        self.attention_dropout = attention_dropout
+        self.classifier_dropout_rate = classifier_dropout_rate
+        self.tie_word_embeddings = tie_word_embeddings
+
+    def __setattr__(self, key, value):
+        shared_attr_with_submodules = [
+            "output_hidden_states",
+            "output_attentions",
+            "_attn_implementation",
+            "dropout_rate",
+            "attention_dropout",
+        ]
+
+        if key in shared_attr_with_submodules:
+            setattr(self.encoder, key, value)
+            setattr(self.decoder, key, value)
+        super().__setattr__(key, value)
+
+    def get_text_config(self, decoder=False) -> "PretrainedConfig":
+        # Always return self, regardless of the decoder option.
+        del decoder
+        return self
+
+
+__all__ = ["T5GemmaConfig", "T5GemmaModuleConfig"]