Add DBRX Model (#29921)

* wip

* fix __init__.py

* add docs

* Apply suggestions from code review
Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com>

* address comments 1

* work on make fixup

* pass configs down

* add sdpa attention

* remove DbrxBlock

* add to configuration_auto

* docstring now passes formatting test

* fix style

* update READMEs

* add dbrx to modeling_auto

* make fix-copies generated this

* add DBRX_PRETRAINED_CONFIG_ARCHIVE_MAP

* config docstring passes formatting test

* rename moe_loss_weight to router_aux_loss_coef

* add to flash-attn documentation

* fix model-path in tests

* Explicitly make `"suli"` the default `ffn_act_fn`
Co-authored-by: Wing Lian <wing.lian@gmail.com>

* default to using router_aux_loss_coef over ffn_config[moe_loss_weight]

* fix _flash_attn_uses_top_left_mask and is_causal

* fix tests path

* don't use token type IDs

* follow Llama and remove token_type_ids from test

* init ConfigTester differently so tests pass

* remove multiple choice test

* remove question + answer test

* remove sequence classification test

* remove token classification test

* copy Llama tests and remove token_type_ids from test inputs

* do not test pruning or headmasking; style code

* add _tied_weights_keys parameter to pass test

* add type hints

* fix type check

* update config tester

* remove masked_lm test

* remove encoder tests

* initialize DbrxModelTester with correct params

* style

* torch_dtype does not rely on torch

* run make fixup, fix-copies

* use https://huggingface.co/v2ray/dbrx-base-fixed/blob/main/modeling_dbrx.py



* add copyright info

* fix imports and DbrxRotaryEmbedding

* update DbrxModel docstring

* use copies

* change model path in docstring

* use config in DbrxFFN

* fix flashattention2, sdpaattention

* input config to DbrXAttention, DbrxNormAttentionNorm

* more fixes

* fix

* fix again!

* add informative comment

* fix ruff?

* remove print statement + style

* change doc-test

* fix doc-test

* fix docstring

* delete commented out text

* make defaults match dbrx-instruct

* replace `router_aux_loss_coef` with `moe_loss_weight`

* is_decoder=True

* remove is_decoder from configtester

* implement sdpa properly

* make is_decoder pass tests

* start on the GenerationTesterMixin tests

* add dbrx to sdpa documentation

* skip weight typing test

* style

* initialize smaller model
Co-authored-by: Matt <Rocketknight1@users.noreply.github.com>

* Add DBRX to toctree

* skip test_new_cache_format

* make config defaults smaller again

* add pad_token_id

* remove pad_token_id from config

* Remove all references to DBRX_PRETRAINED_CONFIG_ARCHIVE_MAP

* Update src/transformers/models/dbrx/__init__.py
Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com>

* Update src/transformers/models/dbrx/modeling_dbrx.py
Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com>

* Update docs/source/en/model_doc/dbrx.md
Co-authored-by: Matt <Rocketknight1@users.noreply.github.com>

* Update src/transformers/models/dbrx/configuration_dbrx.py
Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com>

* Update docs/source/en/model_doc/dbrx.md
Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com>

* fix typo

* Apply suggestions from code review
Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com>

* update docs, fix configuration_auto.py

* address pr comments

* remove is_decoder flag

* slice

* fix requires grad

* remove grad

* disconnect differently

* remove grad

* enable grads

* patch

* detach expert

* nissan al ghaib

* Update modeling_dbrx.py

* Update src/transformers/models/dbrx/modeling_dbrx.py
Co-authored-by: Matt <Rocketknight1@users.noreply.github.com>

* replace "Gemma" with "Dbrx"

* remove # type: ignore

* don't hardcode vocab_size

* remove ToDo

* Re-add removed idefics2 line

* Update test to use tiny-random!

* Remove TODO

* Remove one more case of loading the entire dbrx-instruct in the tests

* Update src/transformers/models/dbrx/modeling_dbrx.py
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* address some comments

* small model

* add dbrx to tokenization_auto

* More docstrings with add_start_docstrings

* Dbrx for now

* add PipelineTesterMixin

* Update src/transformers/models/dbrx/configuration_dbrx.py
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* remove flash-attn2 import error

* fix docstring
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* add useage example

* put on one line
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* fix ffn_act_fn
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* change "dbrx" to "DBRX" for display purposes.

* fix __init__.py?

* fix __init__.py

* fix README

* return the aux_loss

* remove extra spaces

* fix configuration_auto.py

* fix format in tokenization_auto

* remove new line

* add more useage examples

---------
Co-authored-by: Abhi Venigalla <abhi.venigalla@databricks.com>
Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com>
Co-authored-by: Eitan Turok <eitan.turok@databricks.com>
Co-authored-by: Eitan Turok <150733043+eitanturok@users.noreply.github.com>
Co-authored-by: Wing Lian <wing.lian@gmail.com>
Co-authored-by: Eitan Turok <eitanturok@gmail.com>
Co-authored-by: Matt <Rocketknight1@users.noreply.github.com>
Co-authored-by: Matt <rocketknight1@gmail.com>
Co-authored-by: Your Name <you@example.com>
Co-authored-by: Mihir Patel <mihir.v.patel7@gmail.com>
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

src/transformers/models/auto/configuration_auto.py

@@ -77,6 +77,7 @@ CONFIG_MAPPING_NAMES = OrderedDict(
         ("data2vec-audio", "Data2VecAudioConfig"),
         ("data2vec-text", "Data2VecTextConfig"),
         ("data2vec-vision", "Data2VecVisionConfig"),
+        ("dbrx", "DbrxConfig"),
         ("deberta", "DebertaConfig"),
         ("deberta-v2", "DebertaV2Config"),
         ("decision_transformer", "DecisionTransformerConfig"),
@@ -340,6 +341,7 @@ MODEL_NAMES_MAPPING = OrderedDict(
         ("data2vec-audio", "Data2VecAudio"),
         ("data2vec-text", "Data2VecText"),
         ("data2vec-vision", "Data2VecVision"),
+        ("dbrx", "DBRX"),
         ("deberta", "DeBERTa"),
         ("deberta-v2", "DeBERTa-v2"),
         ("decision_transformer", "Decision Transformer"),

src/transformers/models/auto/modeling_auto.py

@@ -75,6 +75,7 @@ MODEL_MAPPING_NAMES = OrderedDict(
         ("data2vec-audio", "Data2VecAudioModel"),
         ("data2vec-text", "Data2VecTextModel"),
         ("data2vec-vision", "Data2VecVisionModel"),
+        ("dbrx", "DbrxModel"),
         ("deberta", "DebertaModel"),
         ("deberta-v2", "DebertaV2Model"),
         ("decision_transformer", "DecisionTransformerModel"),
@@ -439,6 +440,7 @@ MODEL_FOR_CAUSAL_LM_MAPPING_NAMES = OrderedDict(
         ("cpmant", "CpmAntForCausalLM"),
         ("ctrl", "CTRLLMHeadModel"),
         ("data2vec-text", "Data2VecTextForCausalLM"),
+        ("dbrx", "DbrxForCausalLM"),
         ("electra", "ElectraForCausalLM"),
         ("ernie", "ErnieForCausalLM"),
         ("falcon", "FalconForCausalLM"),

src/transformers/models/auto/tokenization_auto.py

@@ -150,6 +150,7 @@ else:
             ("ctrl", ("CTRLTokenizer", None)),
             ("data2vec-audio", ("Wav2Vec2CTCTokenizer", None)),
             ("data2vec-text", ("RobertaTokenizer", "RobertaTokenizerFast" if is_tokenizers_available() else None)),
+            ("dbrx", ("GPT2Tokenizer", "GPT2TokenizerFast" if is_tokenizers_available() else None)),
             ("deberta", ("DebertaTokenizer", "DebertaTokenizerFast" if is_tokenizers_available() else None)),
             (
                 "deberta-v2",

src/transformers/models/dbrx/__init__.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.
+from typing import TYPE_CHECKING
+
+from ...utils import OptionalDependencyNotAvailable, _LazyModule, is_torch_available
+
+
+_import_structure = {
+    "configuration_dbrx": ["DbrxConfig"],
+}
+
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
+    _import_structure["modeling_dbrx"] = [
+        "DbrxForCausalLM",
+        "DbrxModel",
+        "DbrxPreTrainedModel",
+    ]
+
+
+if TYPE_CHECKING:
+    from .configuration_dbrx import DbrxConfig
+
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
+        from .modeling_dbrx import DbrxForCausalLM, DbrxModel, DbrxPreTrainedModel
+
+
+else:
+    import sys
+
+    sys.modules[__name__] = _LazyModule(__name__, globals()["__file__"], _import_structure, module_spec=__spec__)

src/transformers/models/dbrx/configuration_dbrx.py

+# coding=utf-8
+# Copyright 2024 Databricks Mosaic Research 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.
+""" DBRX model configuration """
+
+from typing import Any, Optional
+
+from ...configuration_utils import PretrainedConfig
+from ...utils import logging
+
+
+logger = logging.get_logger(__name__)
+
+
+class DbrxAttentionConfig(PretrainedConfig):
+    """Configuration class for Dbrx Attention.
+
+    [`DbrxAttention`] class. It is used to instantiate attention layers
+    according to the specified arguments, defining the layers architecture.
+
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+    Args:
+        attn_pdrop (`float`, *optional*, defaults to 0.0):
+            The dropout probability for the attention layers.
+        clip_qkv (`float`, *optional*):
+            If set, clip the queries, keys, and values in the attention layer to this value.
+        kv_n_heads (`Optional[int]`, defaults to 1): For grouped_query_attention only, allow user to specify number of kv heads.
+        rope_theta (`float`, defaults to 10000.0): The base frequency for rope.
+    """
+
+    def __init__(
+        self,
+        attn_pdrop: float = 0.0,
+        clip_qkv: Optional[float] = None,
+        kv_n_heads: int = 1,
+        rope_theta: float = 10000.0,
+        **kwargs: Any,
+    ):
+        super().__init__(**kwargs)
+        self.attn_pdrop = attn_pdrop
+        self.clip_qkv = clip_qkv
+        self.kv_n_heads = kv_n_heads
+        self.rope_theta = rope_theta
+
+        for k in ["model_type"]:
+            if k in kwargs:
+                kwargs.pop(k)
+        if len(kwargs) != 0:
+            raise ValueError(f"Found unknown {kwargs=}")
+
+    @classmethod
+    def from_pretrained(cls, pretrained_model_name_or_path: str, **kwargs: Any) -> "PretrainedConfig":
+        cls._set_token_in_kwargs(kwargs)
+
+        config_dict, kwargs = cls.get_config_dict(pretrained_model_name_or_path, **kwargs)
+
+        if config_dict.get("model_type") == "dbrx":
+            config_dict = config_dict["attn_config"]
+
+        if "model_type" in config_dict and hasattr(cls, "model_type") and config_dict["model_type"] != cls.model_type:
+            logger.warning(
+                f"You are using a model of type {config_dict['model_type']} to instantiate a model of type "
+                + f"{cls.model_type}. This is not supported for all configurations of models and can yield errors."
+            )
+
+        return cls.from_dict(config_dict, **kwargs)
+
+
+class DbrxFFNConfig(PretrainedConfig):
+    """Configuration class for Dbrx FFN.
+
+    [`DbrxFFN`] class. It is used to instantiate feedforward layers according to
+    the specified arguments, defining the layers architecture.
+
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+    Args:
+        ffn_act_fn (`dict`, *optional*, defaults to `None`): A dict specifying activation function for the FFN.
+            The dict should have a key 'name' with the value being the name of the activation function along with
+            any additional keyword arguments. If `None`, then set to `{"name": "silu"}`.
+        ffn_hidden_size (`int`, defaults to 3584): The hidden size of the feedforward network.
+        moe_num_experts (`int`, defaults to 4): The number of experts in the mixture of experts layer.
+        moe_top_k (`int`, defaults to 1): The number of experts to use in the mixture of experts layer.
+        moe_jitter_eps (`float`, *optional*, defaults to `None`): If not `None`, the jitter epsilon for the mixture of experts layer.
+        moe_loss_weight (`float`, defaults to 0.01): The loss weight for the mixture of experts layer.
+        moe_normalize_expert_weights (`float`, *optional*, defaults to 1.0): The normalization factor for the expert weights.
+    """
+
+    def __init__(
+        self,
+        ffn_act_fn: dict = None,
+        ffn_hidden_size: int = 3584,
+        moe_num_experts: int = 4,
+        moe_top_k: int = 1,
+        moe_jitter_eps: Optional[float] = None,
+        moe_loss_weight: float = 0.01,
+        moe_normalize_expert_weights: Optional[float] = 1.0,
+        **kwargs: Any,
+    ):
+        super().__init__()
+        if ffn_act_fn is None:
+            ffn_act_fn = {"name": "silu"}
+        self.ffn_act_fn = ffn_act_fn
+        self.ffn_hidden_size = ffn_hidden_size
+        self.moe_num_experts = moe_num_experts
+        self.moe_top_k = moe_top_k
+        self.moe_jitter_eps = moe_jitter_eps
+        self.moe_loss_weight = moe_loss_weight
+        self.moe_normalize_expert_weights = moe_normalize_expert_weights
+
+        for k in ["model_type"]:
+            if k in kwargs:
+                kwargs.pop(k)
+        if len(kwargs) != 0:
+            raise ValueError(f"Found unknown {kwargs=}")
+
+    @classmethod
+    def from_pretrained(cls, pretrained_model_name_or_path: str, **kwargs: Any) -> "PretrainedConfig":
+        cls._set_token_in_kwargs(kwargs)
+
+        config_dict, kwargs = cls.get_config_dict(pretrained_model_name_or_path, **kwargs)
+
+        if config_dict.get("model_type") == "dbrx":
+            config_dict = config_dict["ffn_config"]
+
+        if "model_type" in config_dict and hasattr(cls, "model_type") and config_dict["model_type"] != cls.model_type:
+            logger.warning(
+                f"You are using a model of type {config_dict['model_type']} to instantiate a model of type "
+                + f"{cls.model_type}. This is not supported for all configurations of models and can yield errors."
+            )
+
+        return cls.from_dict(config_dict, **kwargs)
+
+
+class DbrxConfig(PretrainedConfig):
+    r"""
+
+    This is the configuration class to store the configuration of a [`DbrxModel`]. It is used to instantiate a Dbrx model according to the
+    specified arguments, defining the model architecture. Instantiating a configuration with the
+    defaults will yield a different configuration to that of the [databricks/dbrx-instruct](https://huggingface.co/databricks/dbrx-instruct) architecture.
+
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+
+    Args:
+        d_model (`int`, *optional*, defaults to 2048):
+            Dimensionality of the embeddings and hidden states.
+        n_heads (`int`, *optional*, defaults to 16):
+            Number of attention heads for each attention layer in the Transformer encoder.
+        n_layers (`int`, *optional*, defaults to 24):
+            Number of hidden layers in the Transformer encoder.
+        max_seq_len (`int`, *optional*, defaults to 2048):
+            The maximum sequence length of the model.
+        vocab_size (`int`, *optional*, defaults to 32000):
+            Vocabulary size of the Dbrx model. Defines the maximum number of different tokens that can be represented by
+            the `inputs_ids` passed when calling [`DbrxModel`].
+        resid_pdrop (`float`, *optional*, defaults to 0.0):
+            The dropout probability applied to the attention output before combining with residual.
+        emb_pdrop (`float`, *optional*, defaults to 0.0):
+            The dropout probability for the embedding layer.
+        attn_config (`dict`, *optional*):
+            A dictionary used to configure the model's attention module.
+        ffn_config (`dict`, *optional*):
+            A dictionary used to configure the model's FFN module.
+        use_cache (`bool`, *optional*, defaults to `True`):
+            Whether or not the model should return the last key/values attentions (not used by all models).
+        initializer_range (`float`, *optional*, defaults to 0.02):
+            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+        output_router_logits (`bool`, *optional*, defaults to `False`):
+            Whether or not the router logits should be returned by the model. Enabling this will also
+            allow the model to output the auxiliary loss. See [here]() for more details.
+
+
+    Example:
+    ```python
+    >>> from transformers import DbrxConfig, DbrxModel
+
+    >>> # Initializing a Dbrx configuration
+    >>> configuration = DbrxConfig(n_layers=2, d_model=256, n_heads=8, vocab_size=128)
+
+    >>> # Initializing a model (with random weights) from the configuration
+    >>> model = DbrxModel(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```
+    """
+
+    model_type = "dbrx"
+    attribute_map = {
+        "num_attention_heads": "n_heads",
+        "hidden_size": "d_model",
+        "num_hidden_layers": "n_layers",
+        "max_position_embeddings": "max_seq_len",
+    }
+
+    def __init__(
+        self,
+        d_model: int = 2048,
+        n_heads: int = 16,
+        n_layers: int = 24,
+        max_seq_len: int = 2048,
+        vocab_size: int = 32000,
+        resid_pdrop: float = 0.0,
+        emb_pdrop: float = 0.0,
+        attn_config: Optional[DbrxAttentionConfig] = None,
+        ffn_config: Optional[DbrxFFNConfig] = None,
+        use_cache: bool = True,
+        initializer_range: float = 0.02,
+        output_router_logits: bool = False,
+        **kwargs: Any,
+    ):
+        if attn_config is None:
+            self.attn_config = DbrxAttentionConfig()
+        elif isinstance(attn_config, dict):
+            self.attn_config = DbrxAttentionConfig(**attn_config)
+        else:
+            self.attn_config = attn_config
+
+        if ffn_config is None:
+            self.ffn_config = DbrxFFNConfig()
+        elif isinstance(ffn_config, dict):
+            self.ffn_config = DbrxFFNConfig(**ffn_config)
+        else:
+            self.ffn_config = ffn_config
+
+        self.d_model = d_model
+        self.n_heads = n_heads
+        self.n_layers = n_layers
+        self.max_seq_len = max_seq_len
+        self.vocab_size = vocab_size
+        self.resid_pdrop = resid_pdrop
+        self.emb_pdrop = emb_pdrop
+        self.use_cache = use_cache
+        self.initializer_range = initializer_range
+        self.output_router_logits = output_router_logits
+
+        tie_word_embeddings = kwargs.pop("tie_word_embeddings", False)
+        if tie_word_embeddings:
+            raise ValueError("tie_word_embeddings is not supported for DBRX models.")
+
+        super().__init__(tie_word_embeddings=tie_word_embeddings, **kwargs)

src/transformers/models/dbrx/modeling_dbrx.py

+# coding=utf-8
+# Copyright 2024 Databricks Mosaic Research 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.
+""" PyTorch DBRX model. """
+
+import math
+from typing import Any, Dict, Optional, Tuple, Union
+
+import torch
+import torch.nn.functional as F
+import torch.utils.checkpoint
+from torch import nn
+
+from ...activations import ACT2FN
+from ...cache_utils import Cache, DynamicCache, StaticCache
+from ...modeling_attn_mask_utils import AttentionMaskConverter
+from ...modeling_outputs import MoeCausalLMOutputWithPast, MoeModelOutputWithPast
+from ...modeling_utils import PreTrainedModel
+from ...utils import (
+    add_start_docstrings,
+    add_start_docstrings_to_model_forward,
+    is_flash_attn_2_available,
+    is_flash_attn_greater_or_equal_2_10,
+    logging,
+    replace_return_docstrings,
+)
+from .configuration_dbrx import DbrxConfig
+
+
+if is_flash_attn_2_available():
+    from flash_attn import flash_attn_func, flash_attn_varlen_func
+    from flash_attn.bert_padding import index_first_axis, pad_input, unpad_input  # noqa
+
+logger = logging.get_logger(__name__)
+
+_CONFIG_FOR_DOC = "DbrxConfig"
+
+
+# Copied from transformers.models.gemma.modeling_gemma.GemmaRotaryEmbedding with Gemma->Dbrx
+class DbrxRotaryEmbedding(nn.Module):
+    def __init__(self, dim, max_position_embeddings=2048, base=10000, device=None):
+        super().__init__()
+
+        self.dim = dim
+        self.max_position_embeddings = max_position_embeddings
+        self.base = base
+        self.register_buffer("inv_freq", None, persistent=False)
+
+    @torch.no_grad()
+    def forward(self, x, position_ids, seq_len=None):
+        # x: [bs, num_attention_heads, seq_len, head_size]
+        if self.inv_freq is None:
+            self.inv_freq = 1.0 / (
+                self.base ** (torch.arange(0, self.dim, 2, dtype=torch.int64, device=x.device).float() / self.dim)
+            )
+        inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1)
+        position_ids_expanded = position_ids[:, None, :].float()
+        # Force float32 since bfloat16 loses precision on long contexts
+        # See https://github.com/huggingface/transformers/pull/29285
+        device_type = x.device.type
+        device_type = device_type if isinstance(device_type, str) and device_type != "mps" else "cpu"
+        with torch.autocast(device_type=device_type, enabled=False):
+            freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
+            emb = torch.cat((freqs, freqs), dim=-1)
+            cos = emb.cos()
+            sin = emb.sin()
+        return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)
+
+
+# Copied from transformers.models.llama.modeling_llama.rotate_half
+def rotate_half(x):
+    """Rotates half the hidden dims of the input."""
+    x1 = x[..., : x.shape[-1] // 2]
+    x2 = x[..., x.shape[-1] // 2 :]
+    return torch.cat((-x2, x1), dim=-1)
+
+
+# Copied from transformers.models.llama.modeling_llama.apply_rotary_pos_emb
+def apply_rotary_pos_emb(q, k, cos, sin, position_ids=None, unsqueeze_dim=1):
+    """Applies Rotary Position Embedding to the query and key tensors.
+
+    Args:
+        q (`torch.Tensor`): The query tensor.
+        k (`torch.Tensor`): The key tensor.
+        cos (`torch.Tensor`): The cosine part of the rotary embedding.
+        sin (`torch.Tensor`): The sine part of the rotary embedding.
+        position_ids (`torch.Tensor`, *optional*):
+            Deprecated and unused.
+        unsqueeze_dim (`int`, *optional*, defaults to 1):
+            The 'unsqueeze_dim' argument specifies the dimension along which to unsqueeze cos[position_ids] and
+            sin[position_ids] so that they can be properly broadcasted to the dimensions of q and k. For example, note
+            that cos[position_ids] and sin[position_ids] have the shape [batch_size, seq_len, head_dim]. Then, if q and
+            k have the shape [batch_size, heads, seq_len, head_dim], then setting unsqueeze_dim=1 makes
+            cos[position_ids] and sin[position_ids] broadcastable to the shapes of q and k. Similarly, if q and k have
+            the shape [batch_size, seq_len, heads, head_dim], then set unsqueeze_dim=2.
+    Returns:
+        `tuple(torch.Tensor)` comprising of the query and key tensors rotated using the Rotary Position Embedding.
+    """
+    cos = cos.unsqueeze(unsqueeze_dim)
+    sin = sin.unsqueeze(unsqueeze_dim)
+    q_embed = (q * cos) + (rotate_half(q) * sin)
+    k_embed = (k * cos) + (rotate_half(k) * sin)
+    return q_embed, k_embed
+
+
+# Copied from transformers.models.llama.modeling_llama.repeat_kv
+def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
+    """
+    This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch,
+    num_key_value_heads, seqlen, head_dim) to (batch, num_attention_heads, seqlen, head_dim)
+    """
+    batch, num_key_value_heads, slen, head_dim = hidden_states.shape
+    if n_rep == 1:
+        return hidden_states
+    hidden_states = hidden_states[:, :, None, :, :].expand(batch, num_key_value_heads, n_rep, slen, head_dim)
+    return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim)
+
+
+def load_balancing_loss_func(
+    gate_logits: torch.Tensor,
+    num_experts: int,
+    top_k: int,
+    attention_mask: Optional[torch.Tensor],
+) -> torch.Tensor:
+    r"""Computes auxiliary load balancing loss as in Switch Transformer - implemented in Pytorch.
+
+    See Switch Transformer (https://arxiv.org/abs/2101.03961) for more details. This function implements the loss
+    function presented in equations (4) - (6) of the paper. It aims at penalizing cases where the routing between
+    experts is too unbalanced.
+
+    Args:
+        gate_logits (Union[`torch.Tensor`, Tuple[torch.Tensor]):
+            Logits from the `gate`, should be a tuple of model.config.num_hidden_layers tensors of
+            shape [batch_size X sequence_length, num_experts].
+        num_experts (`int`):
+            Number of experts.
+        top_k (`int`):
+            The number of experts each token is routed to.
+        attention_mask (`torch.Tensor`, None):
+            The attention_mask used in forward function
+            shape [batch_size X sequence_length] if not None.
+
+    Returns:
+        The auxiliary loss.
+    """
+    if gate_logits is None or not isinstance(gate_logits, tuple):
+        return torch.tensor(0.0)
+
+    if isinstance(gate_logits, tuple):
+        compute_device = gate_logits[0].device
+        concatenated_gate_logits = torch.cat([layer_gate.to(compute_device) for layer_gate in gate_logits], dim=0)
+
+    routing_weights = torch.nn.functional.softmax(concatenated_gate_logits, dim=-1)
+
+    _, selected_experts = torch.topk(routing_weights, top_k, dim=-1)
+
+    expert_mask = torch.nn.functional.one_hot(selected_experts, num_experts)
+
+    if attention_mask is None:
+        # Compute the percentage of tokens routed to each experts
+        tokens_per_expert = torch.mean(expert_mask.float(), dim=0)
+
+        # Compute the average probability of routing to these experts
+        router_prob_per_expert = torch.mean(routing_weights, dim=0)
+    else:
+        batch_size, sequence_length = attention_mask.shape
+        num_hidden_layers = concatenated_gate_logits.shape[0] // (batch_size * sequence_length)
+
+        # Compute the mask that masks all padding tokens as 0 with the same shape of expert_mask
+        expert_attention_mask = (
+            attention_mask[None, :, :, None, None]
+            .expand((num_hidden_layers, batch_size, sequence_length, top_k, num_experts))
+            .reshape(-1, top_k, num_experts)
+            .to(compute_device)
+        )
+
+        # Compute the percentage of tokens routed to each experts
+        tokens_per_expert = torch.sum(expert_mask.float() * expert_attention_mask, dim=0) / torch.sum(
+            expert_attention_mask, dim=0
+        )
+
+        # Compute the mask that masks all padding tokens as 0 with the same shape of tokens_per_expert
+        router_per_expert_attention_mask = (
+            attention_mask[None, :, :, None]
+            .expand((num_hidden_layers, batch_size, sequence_length, num_experts))
+            .reshape(-1, num_experts)
+            .to(compute_device)
+        )
+
+        # Compute the average probability of routing to these experts
+        router_prob_per_expert = torch.sum(routing_weights * router_per_expert_attention_mask, dim=0) / torch.sum(
+            router_per_expert_attention_mask, dim=0
+        )
+
+    overall_loss = torch.sum(tokens_per_expert * router_prob_per_expert.unsqueeze(0))
+    return overall_loss * num_experts
+
+
+# Copied from transformers.models.llama.modeling_llama._get_unpad_data
+def _get_unpad_data(attention_mask):
+    seqlens_in_batch = attention_mask.sum(dim=-1, dtype=torch.int32)
+    indices = torch.nonzero(attention_mask.flatten(), as_tuple=False).flatten()
+    max_seqlen_in_batch = seqlens_in_batch.max().item()
+    cu_seqlens = F.pad(torch.cumsum(seqlens_in_batch, dim=0, dtype=torch.int32), (1, 0))
+    return (
+        indices,
+        cu_seqlens,
+        max_seqlen_in_batch,
+    )
+
+
+class DbrxAttention(nn.Module):
+    """Multi-head self attention."""
+
+    def __init__(self, config: DbrxConfig, block_idx: Optional[int] = None):
+        super().__init__()
+        self.config = config
+        self.hidden_size = config.d_model
+        self.num_heads = config.n_heads
+        self.head_dim = self.hidden_size // self.num_heads
+        self.max_position_embeddings = config.max_seq_len
+        self.block_idx = block_idx
+        if block_idx is None:
+            logger.warning_once(
+                f"Instantiating {self.__class__.__name__} without passing a `block_idx` is not recommended and will "
+                + "lead to errors during the forward call if caching is used. Please make sure to provide a `block_idx` "
+                + "when creating this class."
+            )
+
+        attn_config = config.attn_config
+        self.attn_pdrop = attn_config.attn_pdrop
+        self.clip_qkv = attn_config.clip_qkv
+        self.num_key_value_heads = attn_config.kv_n_heads
+        self.num_key_value_groups = self.num_heads // self.num_key_value_heads
+        self.rope_theta = attn_config.rope_theta
+        self.is_causal = True
+
+        self.Wqkv = nn.Linear(
+            self.hidden_size, self.hidden_size + 2 * self.num_key_value_heads * self.head_dim, bias=False
+        )
+        self.out_proj = nn.Linear(self.hidden_size, self.hidden_size, bias=False)
+        self.rotary_emb = DbrxRotaryEmbedding(
+            self.head_dim,
+            max_position_embeddings=self.max_position_embeddings,
+            base=self.rope_theta,
+        )
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        position_ids: torch.LongTensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        past_key_value: Optional[Cache] = None,
+        output_attentions: bool = False,
+        use_cache: bool = False,
+        cache_position: Optional[torch.LongTensor] = None,
+        **kwargs: Any,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Cache]]:
+        bsz, q_len, _ = hidden_states.size()
+
+        qkv_states = self.Wqkv(hidden_states)
+        min_val = -self.clip_qkv if self.clip_qkv is not None else None
+        max_val = self.clip_qkv
+        qkv_states = qkv_states.clamp(min=min_val, max=max_val)
+
+        query_states, key_states, value_states = qkv_states.split(
+            [
+                self.hidden_size,
+                self.num_key_value_heads * self.head_dim,
+                self.num_key_value_heads * self.head_dim,
+            ],
+            dim=2,
+        )
+
+        query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
+        key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+        value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+
+        past_key_value = getattr(self, "past_key_value", past_key_value)
+        cos, sin = self.rotary_emb(value_states, position_ids)
+        query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)
+
+        if past_key_value is not None:
+            # sin and cos are specific to RoPE models; position_ids needed for the static cache
+            cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}
+            key_states, value_states = past_key_value.update(key_states, value_states, self.block_idx, cache_kwargs)
+
+        key_states = repeat_kv(key_states, self.num_key_value_groups)
+        value_states = repeat_kv(value_states, self.num_key_value_groups)
+
+        attn_weights = torch.matmul(query_states, key_states.transpose(2, 3)) / math.sqrt(self.head_dim)
+
+        if attention_mask is not None:  # no matter the length, we just slice it
+            causal_mask = attention_mask[:, :, :, : key_states.shape[-2]]
+            attn_weights = attn_weights + causal_mask
+
+        # upcast attention to fp32
+        attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query_states.dtype)
+        attn_weights = nn.functional.dropout(attn_weights, p=self.attn_pdrop, training=self.training)
+        attn_output = torch.matmul(attn_weights, value_states)
+
+        if attn_output.size() != (bsz, self.num_heads, q_len, self.head_dim):
+            raise ValueError(
+                f"`attn_output` should be of size {(bsz, self.num_heads, q_len, self.head_dim)}, but is"
+                + f" {attn_output.size()}"
+            )
+
+        attn_output = attn_output.transpose(1, 2).contiguous()
+        attn_output = attn_output.reshape(bsz, q_len, self.hidden_size)
+        attn_output = self.out_proj(attn_output)
+
+        if not output_attentions:
+            attn_weights = None
+
+        return attn_output, attn_weights, past_key_value
+
+
+class DbrxFlashAttention2(DbrxAttention):
+    """Dbrx flash attention module.
+
+    This module inherits from `DbrxAttention` as the weights of the module stays
+    untouched. The only required change would be on the forward pass where it
+    calls the public API of flash attention.
+    """
+
+    def __init__(self, *args: Any, **kwargs: Any):
+        super().__init__(*args, **kwargs)
+
+        # TODO: Should be removed once Flash Attention for RoCm is bumped to 2.1.
+        # flash_attn<2.1 generates top-left aligned causal mask, while what is needed here is bottom-right alignement, that was made default for flash_attn>=2.1. This attribute is used to handle this difference. Reference: https://github.com/Dao-AILab/flash-attention/releases/tag/v2.1.0.
+        # Beware that with flash_attn<2.1, using q_seqlen != k_seqlen (except for the case q_seqlen == 1) produces a wrong mask (top-left).
+        # From: https://github.com/huggingface/transformers/blob/3b8e2932ce743008f63585aae1e1b8b30dc8b3ac/src/transformers/models/gemma/modeling_gemma.py#L318
+        self._flash_attn_uses_top_left_mask = not is_flash_attn_greater_or_equal_2_10()
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.LongTensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[Cache] = None,
+        output_attentions: bool = False,
+        use_cache: bool = False,
+        cache_position: Optional[torch.LongTensor] = None,
+        **kwargs: Any,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+        logger.info("Implicitly setting `output_attentions` to False as it is not supported in Flash Attention.")
+        output_attentions = False
+
+        bsz, q_len, _ = hidden_states.size()
+
+        qkv_states = self.Wqkv(hidden_states)
+        if self.clip_qkv is not None:
+            qkv_states = qkv_states.clamp(min=-self.clip_qkv, max=self.clip_qkv)
+
+        query_states, key_states, value_states = qkv_states.split(
+            [
+                self.hidden_size,
+                self.num_key_value_heads * self.head_dim,
+                self.num_key_value_heads * self.head_dim,
+            ],
+            dim=2,
+        )
+
+        # Flash attention requires the input to have the shape
+        # batch_size x seq_length x head_dim x hidden_dim
+        # therefore we just need to keep the original shape
+        query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
+        key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+        value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+
+        cos, sin = self.rotary_emb(value_states, position_ids)
+        query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)
+
+        past_key_value = getattr(self, "past_key_value", past_key_value)
+
+        if past_key_value is not None:
+            # sin and cos are specific to RoPE models; cache_position needed for the static cache
+            cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}
+            key_states, value_states = past_key_value.update(key_states, value_states, self.block_idx, cache_kwargs)
+
+        # TODO: These transpose are quite inefficient but Flash Attention requires the layout
+        # [batch_size, sequence_length, num_heads, head_dim]. We would need to refactor the KV cache
+        # to be able to avoid many of these transpose/reshape/view.
+        query_states = query_states.transpose(1, 2)
+        key_states = key_states.transpose(1, 2)
+        value_states = value_states.transpose(1, 2)
+
+        dropout_rate = self.attn_pdrop if self.training else 0.0
+
+        # In PEFT, usually we cast the layer norms in float32 for training stability reasons
+        # therefore the input hidden states gets silently casted in float32. Hence, we need
+        # cast them back in the correct dtype just to be sure everything works as expected.
+        # This might slowdown training & inference so it is recommended to not cast the LayerNorms
+        # in fp32. (LlamaRMSNorm handles it correctly)
+        input_dtype = query_states.dtype
+        if input_dtype == torch.float32:
+            if torch.is_autocast_enabled():
+                target_dtype = torch.get_autocast_gpu_dtype()
+            # Handle the case where the model is quantized
+            elif hasattr(self.config, "_pre_quantization_dtype"):
+                target_dtype = self.config._pre_quantization_dtype
+            else:
+                target_dtype = query_states.dtype
+
+            logger.warning_once(
+                "The input hidden states seems to be silently casted in float32, this might be "
+                + "related to the fact you have upcasted embedding or layer norm layers in "
+                + f"float32. We will cast back the input in {target_dtype}."
+            )
+
+            query_states = query_states.to(target_dtype)
+            key_states = key_states.to(target_dtype)
+            value_states = value_states.to(target_dtype)
+
+        attn_output = self._flash_attention_forward(
+            query_states,
+            key_states,
+            value_states,
+            attention_mask,
+            q_len,
+            dropout=dropout_rate,
+        )
+
+        attn_output = attn_output.reshape(bsz, q_len, self.hidden_size).contiguous()
+        attn_output = self.out_proj(attn_output)
+
+        if not output_attentions:
+            attn_weights = None
+
+        return attn_output, attn_weights, past_key_value
+
+    # Copied from transformers.models.llama.modeling_llama.LlamaFlashAttention2._flash_attention_forward
+    def _flash_attention_forward(
+        self, query_states, key_states, value_states, attention_mask, query_length, dropout=0.0, softmax_scale=None
+    ):
+        """
+        Calls the forward method of Flash Attention - if the input hidden states contain at least one padding token
+        first unpad the input, then computes the attention scores and pad the final attention scores.
+
+        Args:
+            query_states (`torch.Tensor`):
+                Input query states to be passed to Flash Attention API
+            key_states (`torch.Tensor`):
+                Input key states to be passed to Flash Attention API
+            value_states (`torch.Tensor`):
+                Input value states to be passed to Flash Attention API
+            attention_mask (`torch.Tensor`):
+                The padding mask - corresponds to a tensor of size `(batch_size, seq_len)` where 0 stands for the
+                position of padding tokens and 1 for the position of non-padding tokens.
+            dropout (`float`):
+                Attention dropout
+            softmax_scale (`float`, *optional*):
+                The scaling of QK^T before applying softmax. Default to 1 / sqrt(head_dim)
+        """
+        if not self._flash_attn_uses_top_left_mask:
+            causal = self.is_causal
+        else:
+            # TODO: Remove the `query_length != 1` check once Flash Attention for RoCm is bumped to 2.1. For details, please see the comment in LlamaFlashAttention2 __init__.
+            causal = self.is_causal and query_length != 1
+
+        # Contains at least one padding token in the sequence
+        if attention_mask is not None:
+            batch_size = query_states.shape[0]
+            query_states, key_states, value_states, indices_q, cu_seq_lens, max_seq_lens = self._upad_input(
+                query_states, key_states, value_states, attention_mask, query_length
+            )
+
+            cu_seqlens_q, cu_seqlens_k = cu_seq_lens
+            max_seqlen_in_batch_q, max_seqlen_in_batch_k = max_seq_lens
+
+            attn_output_unpad = flash_attn_varlen_func(
+                query_states,
+                key_states,
+                value_states,
+                cu_seqlens_q=cu_seqlens_q,
+                cu_seqlens_k=cu_seqlens_k,
+                max_seqlen_q=max_seqlen_in_batch_q,
+                max_seqlen_k=max_seqlen_in_batch_k,
+                dropout_p=dropout,
+                softmax_scale=softmax_scale,
+                causal=causal,
+            )
+
+            attn_output = pad_input(attn_output_unpad, indices_q, batch_size, query_length)
+        else:
+            attn_output = flash_attn_func(
+                query_states, key_states, value_states, dropout, softmax_scale=softmax_scale, causal=causal
+            )
+
+        return attn_output
+
+    # Copied from transformers.models.llama.modeling_llama.LlamaFlashAttention2._upad_input
+    def _upad_input(self, query_layer, key_layer, value_layer, attention_mask, query_length):
+        indices_k, cu_seqlens_k, max_seqlen_in_batch_k = _get_unpad_data(attention_mask)
+        batch_size, kv_seq_len, num_key_value_heads, head_dim = key_layer.shape
+
+        key_layer = index_first_axis(
+            key_layer.reshape(batch_size * kv_seq_len, num_key_value_heads, head_dim), indices_k
+        )
+        value_layer = index_first_axis(
+            value_layer.reshape(batch_size * kv_seq_len, num_key_value_heads, head_dim), indices_k
+        )
+        if query_length == kv_seq_len:
+            query_layer = index_first_axis(
+                query_layer.reshape(batch_size * kv_seq_len, self.num_heads, head_dim), indices_k
+            )
+            cu_seqlens_q = cu_seqlens_k
+            max_seqlen_in_batch_q = max_seqlen_in_batch_k
+            indices_q = indices_k
+        elif query_length == 1:
+            max_seqlen_in_batch_q = 1
+            cu_seqlens_q = torch.arange(
+                batch_size + 1, dtype=torch.int32, device=query_layer.device
+            )  # There is a memcpy here, that is very bad.
+            indices_q = cu_seqlens_q[:-1]
+            query_layer = query_layer.squeeze(1)
+        else:
+            # The -q_len: slice assumes left padding.
+            attention_mask = attention_mask[:, -query_length:]
+            query_layer, indices_q, cu_seqlens_q, max_seqlen_in_batch_q = unpad_input(query_layer, attention_mask)
+
+        return (
+            query_layer,
+            key_layer,
+            value_layer,
+            indices_q,
+            (cu_seqlens_q, cu_seqlens_k),
+            (max_seqlen_in_batch_q, max_seqlen_in_batch_k),
+        )
+
+
+class DbrxSdpaAttention(DbrxAttention):
+    """
+    Dbrx attention module using torch.nn.functional.scaled_dot_product_attention. This module inherits from
+    `DbrxAttention` as the weights of the module stays untouched. The only changes are on the forward pass to adapt to
+    SDPA API.
+    """
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[Cache] = None,
+        output_attentions: bool = False,
+        use_cache: bool = False,
+        cache_position: Optional[torch.LongTensor] = None,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+        if output_attentions:
+            # TODO: Improve this warning with e.g. `model.config.attn_implementation = "manual"` once this is implemented.
+            logger.warning_once(
+                "DbrxModel is using DbrxSdpaAttention, but `torch.nn.functional.scaled_dot_product_attention` does not support `output_attentions=True`. Falling back to the manual attention implementation, "
+                'but specifying the manual implementation will be required from Transformers version v5.0.0 onwards. This warning can be removed using the argument `attn_implementation="eager"` when loading the model.'
+            )
+            return super().forward(
+                hidden_states=hidden_states,
+                attention_mask=attention_mask,
+                position_ids=position_ids,
+                past_key_value=past_key_value,
+                output_attentions=output_attentions,
+                use_cache=use_cache,
+                cache_position=cache_position,
+            )
+
+        bsz, q_len, _ = hidden_states.size()
+
+        qkv_states = self.Wqkv(hidden_states)
+        if self.clip_qkv is not None:
+            qkv_states = qkv_states.clamp(min=-self.clip_qkv, max=self.clip_qkv)
+
+        query_states, key_states, value_states = qkv_states.split(
+            [
+                self.hidden_size,
+                self.num_key_value_heads * self.head_dim,
+                self.num_key_value_heads * self.head_dim,
+            ],
+            dim=2,
+        )
+
+        query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
+        key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+        value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+
+        cos, sin = self.rotary_emb(value_states, position_ids, seq_len=None)
+        query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, None)
+
+        past_key_value = getattr(self, "past_key_value", past_key_value)
+
+        if past_key_value is not None:
+            # sin and cos are specific to RoPE models; cache_position needed for the static cache
+            cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}
+            key_states, value_states = past_key_value.update(key_states, value_states, self.block_idx, cache_kwargs)
+
+        key_states = repeat_kv(key_states, self.num_key_value_groups)
+        value_states = repeat_kv(value_states, self.num_key_value_groups)
+
+        causal_mask = attention_mask
+        if attention_mask is not None:
+            causal_mask = causal_mask[:, :, :, : key_states.shape[-2]]
+
+        # SDPA with memory-efficient backend is currently (torch==2.1.2) bugged with non-contiguous inputs with custom attn_mask,
+        # Reference: https://github.com/pytorch/pytorch/issues/112577.
+        if query_states.device.type == "cuda" and causal_mask is not None:
+            query_states = query_states.contiguous()
+            key_states = key_states.contiguous()
+            value_states = value_states.contiguous()
+
+        attn_output = torch.nn.functional.scaled_dot_product_attention(
+            query_states,
+            key_states,
+            value_states,
+            attn_mask=causal_mask,
+            dropout_p=self.attn_pdrop if self.training else 0.0,
+        )
+
+        attn_output = attn_output.transpose(1, 2).contiguous()
+        attn_output = attn_output.view(bsz, q_len, -1)
+
+        attn_output = self.out_proj(attn_output)
+
+        return attn_output, None, past_key_value
+
+
+DBRX_ATTENTION_CLASSES = {
+    "eager": DbrxAttention,
+    "flash_attention_2": DbrxFlashAttention2,
+    "sdpa": DbrxSdpaAttention,
+}
+
+
+class DbrxNormAttentionNorm(nn.Module):
+    def __init__(self, config: DbrxConfig, block_idx: Optional[int] = None):
+        super().__init__()
+        self.block_idx = block_idx
+        self.resid_pdrop = config.resid_pdrop
+        self.norm_1 = nn.LayerNorm(config.d_model, bias=False)
+        self.attn = DBRX_ATTENTION_CLASSES[config._attn_implementation](
+            config=config,
+            block_idx=block_idx,
+        )
+        self.norm_2 = nn.LayerNorm(config.d_model, bias=False)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        position_ids: torch.LongTensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        past_key_value: Optional[Cache] = None,
+        output_attentions: bool = False,
+        use_cache: bool = False,
+        cache_position: Optional[torch.LongTensor] = None,
+        **kwargs: Any,
+    ) -> Tuple[torch.Tensor, torch.Tensor, Optional[torch.Tensor], Optional[Cache]]:
+        residual_states = hidden_states
+        hidden_states = self.norm_1(hidden_states).to(hidden_states.dtype)
+
+        hidden_states, attn_weights, past_key_value = self.attn(
+            hidden_states=hidden_states,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_value=past_key_value,
+            output_attentions=output_attentions,
+            use_cache=use_cache,
+            cache_position=cache_position,
+            **kwargs,
+        )
+
+        hidden_states = nn.functional.dropout(hidden_states, p=self.resid_pdrop, training=self.training)
+        hidden_states = hidden_states + residual_states
+
+        residual_states = hidden_states
+        hidden_states = self.norm_2(hidden_states).to(hidden_states.dtype)
+
+        return residual_states, hidden_states, attn_weights, past_key_value
+
+
+class DbrxRouter(nn.Module):
+    def __init__(
+        self,
+        hidden_size: int,
+        moe_num_experts: int,
+        moe_top_k: int,
+        moe_jitter_eps: Optional[float],
+        moe_normalize_expert_weights: Optional[float],
+    ):
+        super().__init__()
+        self.hidden_size = hidden_size
+        self.moe_num_experts = moe_num_experts
+        self.moe_top_k = moe_top_k
+        self.moe_jitter_eps = moe_jitter_eps
+        self.moe_normalize_expert_weights = moe_normalize_expert_weights
+
+        self.layer = nn.Linear(self.hidden_size, self.moe_num_experts, bias=False)
+
+    def forward(self, hidden_states: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor, torch.LongTensor]:
+        if self.training and self.moe_jitter_eps is not None:
+            hidden_states *= torch.empty_like(hidden_states).uniform_(
+                1.0 - self.moe_jitter_eps, 1.0 + self.moe_jitter_eps
+            )
+        hidden_states = hidden_states.view(-1, hidden_states.shape[-1])
+        weights = self.layer(hidden_states).softmax(dim=-1, dtype=torch.float32)
+        top_weights, top_experts = torch.topk(weights, self.moe_top_k, dim=-1)
+
+        top_weights_scale = (
+            torch.norm(top_weights, p=self.moe_normalize_expert_weights, dim=-1, keepdim=True)
+            if self.moe_normalize_expert_weights is not None
+            else 1.0
+        )
+        top_weights = top_weights / top_weights_scale
+
+        weights = weights.to(hidden_states.dtype)
+        top_weights = top_weights.to(hidden_states.dtype)
+        return weights, top_weights, top_experts
+
+
+class DbrxExpertGLU(nn.Module):
+    def __init__(self, hidden_size: int, ffn_hidden_size: int, moe_num_experts: int, ffn_act_fn: dict):
+        super().__init__()
+        self.hidden_size = hidden_size
+        self.ffn_hidden_size = ffn_hidden_size
+        self.moe_num_experts = moe_num_experts
+
+        self.w1 = nn.Parameter(torch.empty(moe_num_experts * ffn_hidden_size, hidden_size))
+        self.v1 = nn.Parameter(torch.empty(moe_num_experts * ffn_hidden_size, hidden_size))
+        self.w2 = nn.Parameter(torch.empty(moe_num_experts * ffn_hidden_size, hidden_size))
+
+        act_fn_name = ffn_act_fn.get("name", "silu")
+        self.activation_fn = ACT2FN[act_fn_name]
+
+    def forward(
+        self, x: torch.Tensor, expert_w1: torch.Tensor, expert_v1: torch.Tensor, expert_w2: torch.Tensor
+    ) -> torch.Tensor:
+        gate_proj = x.matmul(expert_w1.t())
+        up_proj = x.matmul(expert_v1.t())
+        gate_proj = self.activation_fn(gate_proj)
+        intermediate_states = gate_proj * up_proj
+        down_proj = intermediate_states.matmul(expert_w2)
+        return down_proj
+
+
+class DbrxExperts(nn.Module):
+    def __init__(self, hidden_size: int, ffn_hidden_size: int, moe_num_experts: int, ffn_act_fn: dict):
+        super().__init__()
+        self.moe_num_experts = moe_num_experts
+        self.mlp = DbrxExpertGLU(
+            hidden_size=hidden_size,
+            ffn_hidden_size=ffn_hidden_size,
+            moe_num_experts=moe_num_experts,
+            ffn_act_fn=ffn_act_fn,
+        )
+
+    def forward(
+        self, x: torch.Tensor, weights: torch.Tensor, top_weights: torch.Tensor, top_experts: torch.LongTensor
+    ) -> torch.Tensor:
+        bsz, q_len, hidden_size = x.shape
+        x = x.view(-1, hidden_size)
+        out = torch.zeros_like(x)
+
+        expert_mask = nn.functional.one_hot(top_experts, num_classes=self.moe_num_experts).permute(2, 1, 0)
+        # Chunk experts at once to avoid storing full parameter multiple times in autograd
+        w1_chunked = self.mlp.w1.view(self.mlp.moe_num_experts, self.mlp.ffn_hidden_size, self.mlp.hidden_size).chunk(
+            self.moe_num_experts, dim=0
+        )
+        v1_chunked = self.mlp.v1.view(self.mlp.moe_num_experts, self.mlp.ffn_hidden_size, self.mlp.hidden_size).chunk(
+            self.moe_num_experts, dim=0
+        )
+        w2_chunked = self.mlp.w2.view(self.mlp.moe_num_experts, self.mlp.ffn_hidden_size, self.mlp.hidden_size).chunk(
+            self.moe_num_experts, dim=0
+        )
+        w1_chunked = [w1.squeeze(dim=0) for w1 in w1_chunked]
+        v1_chunked = [v1.squeeze(dim=0) for v1 in v1_chunked]
+        w2_chunked = [w2.squeeze(dim=0) for w2 in w2_chunked]
+        for expert_idx in range(0, self.moe_num_experts):
+            topk_idx, token_idx = torch.where(expert_mask[expert_idx])
+            if token_idx.shape[0] == 0:
+                continue
+
+            token_list = token_idx
+            topk_list = topk_idx
+
+            expert_tokens = x[None, token_list].reshape(-1, hidden_size)
+            expert_out = (
+                self.mlp(expert_tokens, w1_chunked[expert_idx], v1_chunked[expert_idx], w2_chunked[expert_idx])
+                * top_weights[token_list, topk_list, None]
+            )
+
+            out.index_add_(0, token_idx, expert_out)
+
+        out = out.reshape(bsz, q_len, hidden_size)
+        return out
+
+
+class DbrxFFN(nn.Module):
+    def __init__(self, config: DbrxConfig):
+        super().__init__()
+
+        ffn_config = config.ffn_config
+        self.router = DbrxRouter(
+            hidden_size=config.d_model,
+            moe_num_experts=ffn_config.moe_num_experts,
+            moe_top_k=ffn_config.moe_top_k,
+            moe_jitter_eps=ffn_config.moe_jitter_eps,
+            moe_normalize_expert_weights=ffn_config.moe_normalize_expert_weights,
+        )
+
+        self.experts = DbrxExperts(
+            hidden_size=config.d_model,
+            ffn_hidden_size=ffn_config.ffn_hidden_size,
+            moe_num_experts=ffn_config.moe_num_experts,
+            ffn_act_fn=ffn_config.ffn_act_fn,
+        )
+
+    def forward(self, x: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
+        weights, top_weights, top_experts = self.router(x)
+        out = self.experts(x, weights, top_weights, top_experts)
+        return out, weights
+
+
+class DbrxBlock(nn.Module):
+    def __init__(self, config: DbrxConfig, block_idx: int):
+        super().__init__()
+        self.hidden_size = config.d_model
+        self.resid_pdrop = config.resid_pdrop
+        self.block_idx = block_idx
+        self.norm_attn_norm = DbrxNormAttentionNorm(
+            config=config,
+            block_idx=block_idx,
+        )
+        self.ffn = DbrxFFN(config=config)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: torch.LongTensor = None,
+        past_key_value: Optional[Cache] = None,
+        output_attentions: Optional[bool] = False,
+        output_router_logits: Optional[bool] = False,
+        use_cache: Optional[bool] = False,
+        cache_position: Optional[torch.LongTensor] = None,
+        **kwargs: Any,
+    ) -> Union[
+        Tuple[torch.Tensor],
+        Tuple[torch.Tensor, Optional[torch.Tensor]],
+        Tuple[torch.Tensor, Optional[Cache]],
+        Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Cache]],
+        Tuple[torch.Tensor, Optional[torch.Tensor], Optional[torch.Tensor]],
+        Tuple[torch.Tensor, Optional[Cache], Optional[torch.Tensor]],
+        Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Cache], Optional[torch.Tensor]],
+    ]:
+        """Forward function for DbrxBlock.
+
+        Args:
+            hidden_states (`torch.Tensor`): input to the layer of shape `(batch, seq_len, embed_dim)`
+            position_ids (`torch.LongTensor`): position ids of shape `(batch, seq_len)`
+            attention_mask (`torch.Tensor`, optional): attention mask of size (batch_size, sequence_length)
+                if flash attention is used or (batch_size, 1, query_sequence_length, key_sequence_length)
+                if default attention is used.
+            past_key_value (`Tuple(torch.Tensor)`, optional): cached past key and value projection states
+            output_attentions (`bool`, optional): Whether or not to return the attentions tensors of all
+                attention layers. See `attentions` under returned tensors for more detail.
+            output_router_logits (`bool`, optional): Whether or not to return the router logits.
+            use_cache (`bool`, optional): If set to `True`, `past_key_values` key value states are
+                returned and can be used to speed up decoding (see `past_key_values`).
+            cache_position (`torch.LongTensor`, optional): position ids of the cache
+        """
+
+        # Norm + Attention + Norm
+        resid_states, hidden_states, self_attn_weights, present_key_value = self.norm_attn_norm(
+            hidden_states=hidden_states,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_value=past_key_value,
+            output_attentions=output_attentions,
+            use_cache=use_cache,
+            cache_position=cache_position,
+            **kwargs,
+        )
+
+        # Fully Connected
+        hidden_states, router_logits = self.ffn(hidden_states)
+        hidden_states = nn.functional.dropout(hidden_states, p=self.resid_pdrop, training=self.training)
+        hidden_states = resid_states + hidden_states
+
+        outputs = (hidden_states,)
+
+        if output_attentions:
+            outputs += (self_attn_weights,)
+
+        if use_cache:
+            outputs += (present_key_value,)
+
+        if output_router_logits:
+            outputs += (router_logits,)
+
+        return outputs
+
+
+DBRX_START_DOCSTRING = r"""
+    This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
+    library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
+    etc.)
+
+    This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.
+    Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
+    and behavior.
+
+    Parameters:
+        config ([`DbrxConfig`]):
+            Model configuration class with all the parameters of the model. Initializing with a config file does not
+            load the weights associated with the model, only the configuration. Check out the
+            [`~PreTrainedModel.from_pretrained`] method to load the model weights.
+"""
+
+
+@add_start_docstrings(
+    "The bare DBRX Model outputting raw hidden-states without any specific head on top.",
+    DBRX_START_DOCSTRING,
+)
+class DbrxPreTrainedModel(PreTrainedModel):
+    config_class = DbrxConfig
+    base_model_prefix = "transformer"
+    supports_gradient_checkpointing = True
+    _no_split_modules = ["DbrxBlock"]
+    _skip_keys_device_placement = ["past_key_values"]
+    _supports_flash_attn_2 = True
+    _supports_sdpa = True
+    _supports_cache_class = True
+
+    def _init_weights(self, module: nn.Module):
+        std = self.config.initializer_range
+        if isinstance(module, nn.Linear):
+            module.weight.data.normal_(mean=0.0, std=std)
+            if module.bias is not None:
+                module.bias.data.zero_()
+        elif isinstance(module, nn.Embedding):
+            module.weight.data.normal_(mean=0.0, std=std)
+            if module.padding_idx is not None:
+                module.weight.data[module.padding_idx].zero_()
+        elif isinstance(module, nn.LayerNorm):
+            module.weight.data.normal_(mean=0.0, std=std)
+            if module.bias is not None:
+                module.bias.data.zero_()
+        elif isinstance(module, DbrxExpertGLU):
+            module.w1.data.normal_(mean=0.0, std=std)
+            module.v1.data.normal_(mean=0.0, std=std)
+            module.w2.data.normal_(mean=0.0, std=std)
+
+    def _setup_cache(self, cache_cls: Any, max_batch_size: int, max_cache_len: int):
+        if self.config._attn_implementation == "flash_attention_2" and cache_cls == StaticCache:
+            raise ValueError(
+                "`static` cache implementation is not compatible with "
+                + "`attn_implementation==flash_attention_2`. Make sure to use "
+                + "`spda` in the mean time and open an issue at https://github.com/huggingface/transformers."
+            )
+
+        for block in self.transformer.blocks:
+            device = block.norm_attn_norm.norm_1.weight.device
+            if hasattr(self.config, "_pre_quantization_dtype"):
+                dtype = self.config._pre_quantization_dtype
+            else:
+                dtype = block.norm_attn_norm.attn.out_proj.weight.dtype
+            block.norm_attn_norm.attn.past_key_value = cache_cls(
+                self.config, max_batch_size, max_cache_len, device=device, dtype=dtype
+            )
+
+    def _reset_cache(self):
+        for block in self.transformer.blocks:
+            block.norm_attn_norm.attn.past_key_value = None
+
+
+DBRX_INPUTS_DOCSTRING = r"""
+    Args:
+        input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
+            Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide
+            it.
+
+            Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
+            [`PreTrainedTokenizer.__call__`] for details.
+
+            [What are input IDs?](../glossary#input-ids)
+        attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
+
+            - 1 for tokens that are **not masked**,
+            - 0 for tokens that are **masked**.
+
+            [What are attention masks?](../glossary#attention-mask)
+
+            Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
+            [`PreTrainedTokenizer.__call__`] for details.
+
+            If `past_key_values` is used, optionally only the last `decoder_input_ids` have to be input (see
+            `past_key_values`).
+
+            If you want to change padding behavior, you should read [`modeling_opt._prepare_decoder_attention_mask`]
+            and modify to your needs. See diagram 1 in [the paper](https://arxiv.org/abs/1910.13461) for more
+            information on the default strategy.
+
+            - 1 indicates the head is **not masked**,
+            - 0 indicates the head is **masked**.
+        position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
+            config.n_positions - 1]`.
+
+            [What are position IDs?](../glossary#position-ids)
+        past_key_values (`Cache` or `tuple(tuple(torch.FloatTensor))`, *optional*):
+            Pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention
+            blocks) that can be used to speed up sequential decoding. This typically consists in the `past_key_values`
+            returned by the model at a previous stage of decoding, when `use_cache=True` or `config.use_cache=True`.
+
+            Two formats are allowed:
+            - a [`~cache_utils.Cache`] instance;
+            - Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of
+            shape `(batch_size, num_heads, sequence_length, embed_size_per_head)`). This is also known as the legacy
+            cache format.
+
+            The model will output the same cache format that is fed as input. If no `past_key_values` are passed, the
+            legacy cache format will be returned.
+
+            If `past_key_values` are used, the user can optionally input only the last `input_ids` (those that don't
+            have their past key value states given to this model) of shape `(batch_size, 1)` instead of all `input_ids`
+            of shape `(batch_size, sequence_length)`.
+        inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
+            Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This
+            is useful if you want more control over how to convert `input_ids` indices into associated vectors than the
+            model's internal embedding lookup matrix.
+        use_cache (`bool`, *optional*):
+            If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see
+            `past_key_values`).
+        output_attentions (`bool`, *optional*):
+            Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
+            tensors for more detail.
+        output_hidden_states (`bool`, *optional*):
+            Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
+            more detail.
+        output_router_logits (`bool`, *optional*):
+            Whether or not to return the logits of all the routers. They are useful for computing the router loss, and
+            should not be returned during inference.
+        return_dict (`bool`, *optional*):
+            Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
+        cache_position (`torch.LongTensor` of shape `(sequence_length)`, *optional*):
+            Indices depicting the position of the input sequence tokens in the sequence. Contrarily to `position_ids`,
+            this tensor is not affected by padding. It is used to update the cache in the correct position and to infer
+            the complete sequence length.
+"""
+
+
+@add_start_docstrings(
+    "The bare DBRX Model outputting raw hidden-states without any specific head on top.",
+    DBRX_START_DOCSTRING,
+)
+class DbrxModel(DbrxPreTrainedModel):
+    """Transformer decoder consisting of *config.num_hidden_layers*. Each layer is a [`DbrxBlock`] layer.
+
+    Args:
+        config ([`DbrxConfig`]): Model configuration class with all parameters of the model.
+            Initializing with a config file does not load the weights associated with the model, only the
+            configuration. Check out the [`~PreTrainedModel.from_pretrained`] method to load the model weights.
+    """
+
+    def __init__(self, config: DbrxConfig):
+        super().__init__(config)
+        self.padding_idx = config.pad_token_id
+        self.vocab_size = config.vocab_size
+        self.emb_pdrop = config.emb_pdrop
+
+        self.wte = nn.Embedding(config.vocab_size, config.d_model, self.padding_idx)
+        self.blocks = nn.ModuleList([DbrxBlock(config, block_idx) for block_idx in range(config.n_layers)])
+        self.norm_f = nn.LayerNorm(config.d_model, bias=False)
+        self.gradient_checkpointing = False
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self) -> nn.Embedding:
+        return self.wte
+
+    def set_input_embeddings(self, value: nn.Embedding):
+        self.wte = value
+
+    @add_start_docstrings_to_model_forward(DBRX_INPUTS_DOCSTRING)
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[Cache] = None,
+        inputs_embeds: Optional[torch.Tensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        output_router_logits: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+    ) -> Union[Tuple, MoeModelOutputWithPast]:
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        output_router_logits = (
+            output_router_logits if output_router_logits is not None else self.config.output_router_logits
+        )
+        use_cache = use_cache if use_cache is not None else self.config.use_cache
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        if (input_ids is None) ^ (inputs_embeds is not None):
+            raise ValueError(
+                "You cannot specify both input_ids and inputs_embeds at the same time, and must specify either one"
+            )
+
+        if self.gradient_checkpointing and self.training and use_cache:
+            logger.warning_once(
+                "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`."
+            )
+            use_cache = False
+
+        if inputs_embeds is None:
+            inputs_embeds = self.wte(input_ids)
+
+        inputs_embeds = nn.functional.dropout(inputs_embeds, p=self.emb_pdrop, training=self.training)
+
+        past_seen_tokens = 0
+        if use_cache:  # kept for BC (cache positions)
+            if not isinstance(past_key_values, StaticCache):
+                past_key_values = DynamicCache.from_legacy_cache(past_key_values)
+                past_seen_tokens = past_key_values.get_seq_length()
+
+        if cache_position is None:
+            if isinstance(past_key_values, StaticCache):
+                raise ValueError("cache_position is a required argument when using StaticCache.")
+            cache_position = torch.arange(
+                past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device
+            )
+
+        if position_ids is None:
+            position_ids = cache_position.unsqueeze(0)
+        causal_mask = self._update_causal_mask(attention_mask, inputs_embeds, cache_position)
+
+        # embed positions
+        hidden_states = inputs_embeds
+
+        # decoder layers
+        all_hidden_states = () if output_hidden_states else None
+        all_self_attns = () if output_attentions else None
+        all_router_logits = () if output_router_logits else None
+        next_decoder_cache = None
+
+        for block in self.blocks:
+            if output_hidden_states:
+                all_hidden_states += (hidden_states,)
+
+            if self.gradient_checkpointing and self.training:
+                block_outputs = self._gradient_checkpointing_func(
+                    block.__call__,
+                    hidden_states,
+                    causal_mask,
+                    position_ids,
+                    past_key_values,
+                    output_attentions,
+                    output_router_logits,
+                    use_cache,
+                    cache_position,
+                )
+            else:
+                block_outputs = block(
+                    hidden_states,
+                    attention_mask=causal_mask,
+                    position_ids=position_ids,
+                    past_key_value=past_key_values,
+                    output_attentions=output_attentions,
+                    output_router_logits=output_router_logits,
+                    use_cache=use_cache,
+                    cache_position=cache_position,
+                )
+
+            hidden_states = block_outputs[0]
+
+            if use_cache:
+                next_decoder_cache = block_outputs[2 if output_attentions else 1]
+
+            if output_attentions:
+                all_self_attns += (block_outputs[1],)
+
+            if output_router_logits:
+                all_router_logits += (block_outputs[-1],)
+
+        hidden_states = self.norm_f(hidden_states)
+
+        # add hidden states from the last decoder layer
+        if output_hidden_states:
+            all_hidden_states += (hidden_states,)
+
+        next_cache = None
+        if use_cache:
+            next_cache = (
+                next_decoder_cache.to_legacy_cache() if isinstance(next_decoder_cache, Cache) else next_decoder_cache
+            )
+        if not return_dict:
+            return tuple(
+                v
+                for v in [hidden_states, next_cache, all_hidden_states, all_self_attns, all_router_logits]
+                if v is not None
+            )
+        return MoeModelOutputWithPast(
+            last_hidden_state=hidden_states,
+            past_key_values=next_cache,
+            hidden_states=all_hidden_states,
+            attentions=all_self_attns,
+            router_logits=all_router_logits,
+        )
+
+    # TODO: As of torch==2.2.0, the `attention_mask` passed to the model in `generate` is 2D and of dynamic length even when the static
+    # KV cache is used. This is an issue for torch.compile which then recaptures cudagraphs at each decode steps due to the dynamic shapes.
+    # (`recording cudagraph tree for symint key 13`, etc.), which is VERY slow. A workaround is `@torch.compiler.disable`, but this prevents using
+    # `fullgraph=True`. See more context in https://github.com/huggingface/transformers/pull/29114
+    def _update_causal_mask(
+        self, attention_mask: Optional[torch.Tensor], input_tensor: torch.Tensor, cache_position: torch.Tensor
+    ) -> Optional[torch.Tensor]:
+        if self.config._attn_implementation == "flash_attention_2":
+            if attention_mask is not None and 0.0 in attention_mask:
+                return attention_mask
+            return None
+
+        dtype, device = input_tensor.dtype, input_tensor.device
+        min_dtype = torch.finfo(dtype).min
+        sequence_length = input_tensor.shape[1]
+        if hasattr(self.blocks[0].norm_attn_norm.attn, "past_key_value"):  # static cache
+            target_length = self.config.max_position_embeddings
+        else:  # dynamic cache
+            target_length = (
+                attention_mask.shape[-1] if isinstance(attention_mask, torch.Tensor) else cache_position[-1] + 1
+            )
+        target_length = int(target_length)
+
+        causal_mask = torch.full((sequence_length, target_length), fill_value=min_dtype, dtype=dtype, device=device)
+        if sequence_length != 1:
+            causal_mask = torch.triu(causal_mask, diagonal=1)
+        causal_mask *= torch.arange(target_length, device=device) > cache_position.reshape(-1, 1)
+        causal_mask = causal_mask[None, None, :, :].expand(input_tensor.shape[0], 1, -1, -1)
+        if attention_mask is not None:
+            causal_mask = causal_mask.clone()  # copy to contiguous memory for in-place edit
+            if attention_mask.dim() == 2:
+                mask_length = attention_mask.shape[-1]
+                padding_mask = causal_mask[..., :mask_length].eq(0.0) * attention_mask[:, None, None, :].eq(0.0)
+                causal_mask[..., :mask_length] = causal_mask[..., :mask_length].masked_fill(padding_mask, min_dtype)
+            elif attention_mask.dim() == 4:
+                # backwards compatibility: we allow passing a 4D attention mask shorter than the input length with
+                # cache. In that case, the 4D attention mask attends to the newest tokens only.
+                if attention_mask.shape[-2] < cache_position[0] + sequence_length:
+                    offset = cache_position[0]
+                else:
+                    offset = 0
+                mask_shape = attention_mask.shape
+                mask_slice = (attention_mask.eq(0.0)).to(dtype=dtype) * min_dtype
+                causal_mask[
+                    : mask_shape[0], : mask_shape[1], offset : mask_shape[2] + offset, : mask_shape[3]
+                ] = mask_slice
+
+        if (
+            self.config._attn_implementation == "sdpa"
+            and attention_mask is not None
+            and attention_mask.device.type == "cuda"
+        ):
+            # TODO: For dynamo, rather use a check on fullgraph=True once this is possible (https://github.com/pytorch/pytorch/pull/120400).
+            is_tracing = (
+                torch.jit.is_tracing()
+                or isinstance(input_tensor, torch.fx.Proxy)
+                or (hasattr(torch, "_dynamo") and torch._dynamo.is_compiling())
+            )
+            if not is_tracing and torch.any(attention_mask != 1):
+                # Attend to all tokens in fully masked rows in the causal_mask, for example the relevant first rows when
+                # using left padding. This is required by F.scaled_dot_product_attention memory-efficient attention path.
+                # Details: https://github.com/pytorch/pytorch/issues/110213
+                causal_mask = AttentionMaskConverter._unmask_unattended(causal_mask, min_dtype)
+
+        return causal_mask
+
+
+@add_start_docstrings("The DBRX Model transformer for causal language modeling.", DBRX_START_DOCSTRING)
+class DbrxForCausalLM(DbrxPreTrainedModel):
+    def __init__(self, config: DbrxConfig):
+        super().__init__(config)
+        self.transformer = DbrxModel(config)
+        self.vocab_size = config.vocab_size
+        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+        self.moe_loss_weight = config.ffn_config.moe_loss_weight
+        self.num_experts = config.ffn_config.moe_num_experts
+        self.num_experts_per_tok = config.ffn_config.moe_top_k
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self) -> nn.Embedding:
+        return self.transformer.get_input_embeddings()
+
+    def set_input_embeddings(self, value: nn.Embedding):
+        self.transformer.set_input_embeddings(value)
+
+    def get_output_embeddings(self) -> nn.Linear:
+        return self.lm_head
+
+    def set_output_embeddings(self, new_embeddings: nn.Linear):
+        self.lm_head = new_embeddings
+
+    def set_decoder(self, decoder: DbrxModel):
+        self.transformer = decoder
+
+    def get_decoder(self) -> DbrxModel:
+        return self.transformer
+
+    @add_start_docstrings_to_model_forward(DBRX_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=MoeCausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[Cache] = None,
+        inputs_embeds: Optional[torch.Tensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        output_router_logits: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+    ) -> Union[Tuple, MoeCausalLMOutputWithPast]:
+        r"""Forward function for causal language modeling.
+
+        Args:
+            labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+                Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
+                config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
+                (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
+
+        Returns:
+
+        Example:
+
+        ```python
+        >> from transformers import AutoTokenizer, DbrxForCausalLM
+
+        >> model = DbrxForCausalLM.from_pretrained("databricks/dbrx-instruct")
+        >> tokenizer = AutoTokenizer.from_pretrained("databricks/dbrx-instruct")
+
+        >> prompt = "Hey, are you conscious? Can you talk to me?"
+        >> inputs = tokenizer(prompt, return_tensors="pt")
+
+        >> # Generate
+        >> generate_ids = model.generate(inputs.input_ids, max_length=30)
+        >> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
+        "Hey, are you conscious? Can you talk to me?\nI'm not conscious, but I can talk to you."
+        ```
+        """
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        output_router_logits = (
+            output_router_logits if output_router_logits is not None else self.config.output_router_logits
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
+        outputs = self.transformer(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            output_router_logits=output_router_logits,
+            return_dict=return_dict,
+            cache_position=cache_position,
+        )
+
+        hidden_states = outputs[0]
+        logits = self.lm_head(hidden_states)
+
+        loss = None
+        if labels is not None:
+            # Shift so that tokens < n predict n
+            shift_logits = logits[..., :-1, :].contiguous()
+            shift_labels = labels[..., 1:].contiguous()
+            # Flatten the tokens
+            loss_fct = nn.CrossEntropyLoss()
+            shift_logits = shift_logits.view(-1, self.config.vocab_size)
+            shift_labels = shift_labels.view(-1)
+            # Enable model parallelism
+            shift_labels = shift_labels.to(shift_logits.device)
+            loss = loss_fct(shift_logits, shift_labels)
+
+        aux_loss = None
+        if output_router_logits:
+            aux_loss = load_balancing_loss_func(
+                outputs.router_logits if return_dict else outputs[-1],
+                self.num_experts,
+                self.num_experts_per_tok,
+                attention_mask,
+            )
+            if labels is not None and loss is not None:
+                loss += self.moe_loss_weight * aux_loss.to(loss.device)  # make sure to reside in the same device
+
+        if not return_dict:
+            output = (logits,) + outputs[1:]
+            if output_router_logits:
+                output = (aux_loss,) + output
+            return (loss,) + output if loss is not None else output
+
+        return MoeCausalLMOutputWithPast(
+            loss=loss,
+            aux_loss=aux_loss,
+            logits=logits,
+            past_key_values=outputs.past_key_values,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+            router_logits=outputs.router_logits,
+        )
+
+    def prepare_inputs_for_generation(
+        self,
+        input_ids: torch.Tensor,
+        past_key_values: Optional[Cache] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        inputs_embeds: Optional[torch.Tensor] = None,
+        **kwargs: Any,
+    ) -> Dict[str, Any]:
+        past_length = 0
+        if past_key_values is not None:
+            if isinstance(past_key_values, Cache):
+                cache_length = past_key_values.get_seq_length()
+                past_length = past_key_values.seen_tokens
+                max_cache_length = past_key_values.get_max_length()
+            else:
+                cache_length = past_length = past_key_values[0][0].shape[2]
+                max_cache_length = None
+
+            # Keep only the unprocessed tokens:
+            # 1 - If the length of the attention_mask exceeds the length of input_ids, then we are in a setting where
+            # some of the inputs are exclusively passed as part of the cache (e.g. when passing input_embeds as
+            # input)
+            if attention_mask is not None and attention_mask.shape[1] > input_ids.shape[1]:
+                input_ids = input_ids[:, -(attention_mask.shape[1] - past_length) :]
+            # 2 - If the past_length is smaller than input_ids', then input_ids holds all input tokens. We can discard
+            # input_ids based on the past_length.
+            elif past_length < input_ids.shape[1]:
+                input_ids = input_ids[:, past_length:]
+            # 3 - Otherwise (past_length >= input_ids.shape[1]), let's assume input_ids only has unprocessed tokens.
+
+            # If we are about to go beyond the maximum cache length, we need to crop the input attention mask.
+            if (
+                max_cache_length is not None
+                and attention_mask is not None
+                and cache_length + input_ids.shape[1] > max_cache_length
+            ):
+                attention_mask = attention_mask[:, -max_cache_length:]
+
+        position_ids = kwargs.get("position_ids", None)
+        if attention_mask is not None and position_ids is None:
+            # create position_ids on the fly for batch generation
+            position_ids = attention_mask.long().cumsum(-1) - 1
+            position_ids.masked_fill_(attention_mask == 0, 1)
+            if past_key_values:
+                position_ids = position_ids[:, -input_ids.shape[1] :]
+
+        if self.generation_config.cache_implementation == "static":
+            # generation with static cache
+            cache_position = kwargs.get("cache_position", None)
+            if cache_position is None:
+                past_length = 0
+            else:
+                past_length = cache_position[-1] + 1
+            input_ids = input_ids[:, past_length:]
+            position_ids = position_ids[:, past_length:] if position_ids is not None else None
+
+        # TODO @gante we should only keep a `cache_position` in generate, and do +=1.
+        # same goes for position ids. Could also help with continued generation.
+        input_length = position_ids.shape[-1] if position_ids is not None else input_ids.shape[-1]
+        cache_position = torch.arange(past_length, past_length + input_length, device=input_ids.device)
+        position_ids = position_ids.contiguous() if position_ids is not None else None
+
+        # if `inputs_embeds` are passed, we only want to use them in the 1st generation step
+        if inputs_embeds is not None and past_key_values is None:
+            model_inputs = {"inputs_embeds": inputs_embeds}
+        else:
+            # The `contiguous()` here is necessary to have a static stride during decoding. torchdynamo otherwise
+            # recompiles graphs as the stride of the inputs is a guard. Ref: https://github.com/huggingface/transformers/pull/29114
+            # TODO: use `next_tokens` directly instead.
+            model_inputs = {"input_ids": input_ids.contiguous()}
+
+        model_inputs.update(
+            {
+                "position_ids": position_ids,
+                "cache_position": cache_position,
+                "past_key_values": past_key_values,
+                "use_cache": kwargs.get("use_cache"),
+                "attention_mask": attention_mask,
+            }
+        )
+        return model_inputs
+
+    @staticmethod
+    def _reorder_cache(past_key_values: Cache, beam_idx: torch.LongTensor):
+        reordered_past = ()
+        for layer_past in past_key_values:
+            reordered_past += (
+                tuple(past_state.index_select(0, beam_idx.to(past_state.device)) for past_state in layer_past),
+            )
+        return reordered_past

src/transformers/utils/dummy_pt_objects.py

@@ -2457,6 +2457,27 @@ class Data2VecVisionPreTrainedModel(metaclass=DummyObject):
         requires_backends(self, ["torch"])
 
 
+class DbrxForCausalLM(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
+class DbrxModel(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
+class DbrxPreTrainedModel(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
 DEBERTA_PRETRAINED_MODEL_ARCHIVE_LIST = None
 
 

templates/adding_a_new_model/README.md

@@ -25,7 +25,7 @@ Jump to the [Add new model like section](#add-new-model-like-command) to learn h
 
 ## Cookiecutter Templates
 
-Using the `cookiecutter` utility requires to have all the `dev` dependencies installed. Let's first clone the 
+Using the `cookiecutter` utility requires to have all the `dev` dependencies installed. Let's first clone the
 repository and install it in our environment:
 
 ```shell script
@@ -53,20 +53,20 @@ This should launch the `cookiecutter` package which should prompt you to fill in
 The `modelname` should be cased according to the plain text casing, i.e., BERT, RoBERTa, DeBERTa.
 ```
 modelname [<ModelNAME>]:
-uppercase_modelname [<MODEL_NAME>]: 
-lowercase_modelname [<model_name>]: 
-camelcase_modelname [<ModelName>]: 
+uppercase_modelname [<MODEL_NAME>]:
+lowercase_modelname [<model_name>]:
+camelcase_modelname [<ModelName>]:
 ```
 
 Fill in the `authors` with your team members:
 ```
-authors [The HuggingFace Team]: 
+authors [The HuggingFace Team]:
 ```
 
 The checkpoint identifier is the checkpoint that will be used in the examples across the files. Put the name you wish,
 as it will appear on the modelhub. Do not forget to include the organisation.
 ```
-checkpoint_identifier [organisation/<model_name>-base-cased]: 
+checkpoint_identifier [organisation/<model_name>-base-cased]:
 ```
 
 The tokenizer should either be based on BERT if it behaves exactly like the BERT tokenizer, or a standalone otherwise.
@@ -74,19 +74,19 @@ The tokenizer should either be based on BERT if it behaves exactly like the BERT
 Select tokenizer_type:
 1 - Based on BERT
 2 - Standalone
-Choose from 1, 2 [1]: 
+Choose from 1, 2 [1]:
 ```
 <!---
 Choose if your model is an encoder-decoder, or an encoder-only architecture.
 
-If your model is an encoder-only architecture, the generated architecture will be based on the BERT model. 
+If your model is an encoder-only architecture, the generated architecture will be based on the BERT model.
 If your model is an encoder-decoder architecture, the generated architecture will be based on the BART model. You can,
 of course, edit the files once the generation is complete.
 ```
 Select is_encoder_decoder_model:
 1 - True
 2 - False
-Choose from 1, 2 [1]: 
+Choose from 1, 2 [1]:
 ```
 -->
 
@@ -97,8 +97,8 @@ src/transformers/models/<model_name>/configuration_<model_name>.py
 src/transformers/models/<model_name>/modeling_<model_name>.py
 src/transformers/models/<model_name>/modeling_tf_<model_name>.py
 src/transformers/models/<model_name>/tokenization_<model_name>.py
-tests/test_modeling_<model_name>.py
-tests/test_modeling_tf_<model_name>.py
+tests/models/<model_name>/test_modeling_<model_name>.py
+tests/models/<model_name>/test_modeling_tf_<model_name>.py
 ```
 
 You can run the tests to ensure that they all pass:
@@ -107,9 +107,9 @@ You can run the tests to ensure that they all pass:
 python -m pytest ./tests/test_*<model_name>*.py
 ```
 
-Feel free to modify each file to mimic the behavior of your model. 
+Feel free to modify each file to mimic the behavior of your model.
 
-⚠ You should be careful about the classes preceded by the following line:️ 
+⚠ You should be careful about the classes preceded by the following line:️
 
 ```python
 # Copied from transformers.[...]
@@ -119,8 +119,8 @@ This line ensures that the copy does not diverge from the source. If it *should*
 is different, this line needs to be deleted. If you don't delete this line and run `make fix-copies`,
 your changes will be overwritten.
 
-Once you have edited the files to fit your architecture, simply re-run the tests (and edit them if a change 
-is needed!) afterwards to make sure everything works as expected. 
+Once you have edited the files to fit your architecture, simply re-run the tests (and edit them if a change
+is needed!) afterwards to make sure everything works as expected.
 
 Once the files are generated and you are happy with your changes, here's a checklist to ensure that your contribution
 will be merged quickly:
@@ -251,7 +251,7 @@ Once you're done, you can run the tests to ensure that they all pass:
 python -m pytest ./tests/test_*<model_name>*.py
 ```
 
-⚠ You should be careful about the classes preceded by the following line:️ 
+⚠ You should be careful about the classes preceded by the following line:️
 
 ```python
 # Copied from transformers.[...]
@@ -261,8 +261,8 @@ This line ensures that the copy does not diverge from the source. If it *should*
 is different, this line needs to be deleted. If you don't delete this line and run `make fix-copies`,
 your changes will be overwritten.
 
-Once you have edited the files to fit your architecture, simply re-run the tests (and edit them if a change 
-is needed!) afterwards to make sure everything works as expected. 
+Once you have edited the files to fit your architecture, simply re-run the tests (and edit them if a change
+is needed!) afterwards to make sure everything works as expected.
 
 Once the files are generated and you are happy with your changes, here's a checklist to ensure that your contribution
 will be merged quickly:

tests/models/dbrx/test_modeling_dbrx.py

+# coding=utf-8
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" Testing suite for the PyTorch DBRX model. """
+
+
+import unittest
+
+from parameterized import parameterized
+
+from transformers import DbrxConfig, is_torch_available
+from transformers.testing_utils import require_torch, slow, torch_device
+
+from ...generation.test_utils import GenerationTesterMixin
+from ...test_configuration_common import ConfigTester
+from ...test_modeling_common import ModelTesterMixin, ids_tensor, random_attention_mask
+from ...test_pipeline_mixin import PipelineTesterMixin
+
+
+if is_torch_available():
+    import torch
+
+    from transformers import DbrxForCausalLM, DbrxModel
+
+
+class DbrxModelTester:
+    def __init__(
+        self,
+        parent,
+        hidden_size=32,
+        ffn_hidden_size=32,
+        num_attention_heads=4,
+        kv_n_heads=4,
+        num_hidden_layers=5,
+        max_position_embeddings=512,
+        type_vocab_size=16,
+        batch_size=13,
+        seq_length=7,
+        is_training=True,
+        use_input_mask=True,
+        use_token_type_ids=False,
+        use_labels=True,
+        use_cache=True,
+        type_sequence_label_size=2,
+        num_labels=3,
+        num_choices=4,
+        scope=None,
+        clip_qkv=8,
+        rope_theta=500000,
+        attn_config_model_type="",
+        emb_pdrop=0.0,
+        moe_jitter_eps=0,
+        moe_loss_weight=0.05,
+        moe_num_experts=16,
+        moe_top_k=4,
+        ffn_config_model_type="",
+        ffn_act_fn_name="gelu",
+        initializer_range=0.02,
+        output_router_logits=False,
+        resid_pdrop=0.0,
+        tie_word_embeddings=False,
+        torch_dtype="bfloat16",
+        vocab_size=99,
+        is_decoder=True,
+        pad_token_id=0,
+    ):
+        # Parameters unique to testing
+        self.batch_size = batch_size
+        self.seq_length = seq_length
+        self.type_vocab_size = type_vocab_size
+        self.type_sequence_label_size = type_sequence_label_size
+        self.num_labels = num_labels
+        self.num_choices = num_choices
+        self.scope = scope
+        self.parent = parent
+        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
+
+        # attn_config params
+        self.clip_qkv = clip_qkv
+        self.kv_n_heads = kv_n_heads
+        self.rope_theta = rope_theta
+        self.attn_config_model_type = attn_config_model_type
+
+        # ffn_config params
+        self.ffn_hidden_size = ffn_hidden_size
+        self.moe_jitter_eps = moe_jitter_eps
+        self.moe_loss_weight = moe_loss_weight
+        self.moe_num_experts = moe_num_experts
+        self.moe_top_k = moe_top_k
+        self.ffn_config_model_type = ffn_config_model_type
+        self.ffn_act_fn_name = ffn_act_fn_name
+
+        # Other model params
+        self.hidden_size = hidden_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.max_position_embeddings = max_position_embeddings
+        self.vocab_size = vocab_size
+        self.use_cache = use_cache
+        self.initializer_range = initializer_range
+        self.emb_pdrop = emb_pdrop
+        self.output_router_logits = output_router_logits
+        self.resid_pdrop = resid_pdrop
+        self.tie_word_embeddings = tie_word_embeddings
+        self.torch_dtype = torch_dtype
+        self.is_decoder = is_decoder
+        self.pad_token_id = pad_token_id
+
+        # Make the dictionaries
+        self.ffn_config = {
+            "ffn_hidden_size": self.ffn_hidden_size,
+            "moe_jitter_eps": self.moe_jitter_eps,
+            "moe_loss_weight": self.moe_loss_weight,
+            "moe_num_experts": self.moe_num_experts,
+            "moe_top_k": self.moe_top_k,
+            "model_type": self.ffn_config_model_type,
+            "ffn_act_fn": {"name": self.ffn_act_fn_name},
+        }
+        self.attn_config = {
+            "clip_qkv": self.clip_qkv,
+            "kv_n_heads": self.kv_n_heads,
+            "model_type": self.attn_config_model_type,
+            "rope_theta": self.rope_theta,
+        }
+
+    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])
+
+        token_type_ids = None
+        if self.use_token_type_ids:
+            token_type_ids = ids_tensor([self.batch_size, self.seq_length], self.type_vocab_size)
+
+        sequence_labels = None
+        token_labels = None
+        choice_labels = None
+        if self.use_labels:
+            sequence_labels = ids_tensor([self.batch_size], self.type_sequence_label_size)
+            token_labels = ids_tensor([self.batch_size, self.seq_length], self.num_labels)
+            choice_labels = ids_tensor([self.batch_size], self.num_choices)
+
+        config = self.get_config()
+
+        return config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
+
+    def get_config(self):
+        # Behind the scenes, `DbrxConfig` maps the parameters `hidden_size`, `num_hidden_layers`,
+        # `num_attention_heads`, `max_position_embeddings` to the parameters `d_model`, `n_layers`,
+        # `n_heads`, `max_seq_len` respectively. We use the first group of parameters because
+        # other tests expect every model to have these parameters with these specific names.
+        config = DbrxConfig(
+            vocab_size=self.vocab_size,
+            hidden_size=self.hidden_size,  # mapped to `d_model`
+            num_hidden_layers=self.num_hidden_layers,  # mapped to `n_layers`
+            num_attention_heads=self.num_attention_heads,  # mapped to `n_heads`
+            max_position_embeddings=self.max_position_embeddings,  # mapped to `max_seq_len`
+            attn_config=self.attn_config,
+            ffn_config=self.ffn_config,
+            resid_pdrop=self.resid_pdrop,
+            emb_pdrop=self.emb_pdrop,
+            use_cache=self.use_cache,
+            initializer_range=self.initializer_range,
+            output_router_logits=self.output_router_logits,
+            is_decoder=self.is_decoder,
+            pad_token_id=self.pad_token_id,
+        )
+        return config
+
+    # Copied from tests.models.llama.test_modeling_llama.LlamaModelTester.create_and_check_model with Llama->Dbrx
+    def create_and_check_model(
+        self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
+    ):
+        model = DbrxModel(config=config)
+        model.to(torch_device)
+        model.eval()
+        result = model(input_ids, attention_mask=input_mask)
+        result = model(input_ids)
+        self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
+
+    # Copied from tests.models.llama.test_modeling_llama.LlamaModelTester.create_and_check_model_as_decoder with Llama->Dbrx
+    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 = DbrxModel(config)
+        model.to(torch_device)
+        model.eval()
+        result = model(
+            input_ids,
+            attention_mask=input_mask,
+            encoder_hidden_states=encoder_hidden_states,
+            encoder_attention_mask=encoder_attention_mask,
+        )
+        result = model(
+            input_ids,
+            attention_mask=input_mask,
+            encoder_hidden_states=encoder_hidden_states,
+        )
+        result = model(input_ids, attention_mask=input_mask)
+        self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
+
+    # Copied from tests.models.llama.test_modeling_llama.LlamaModelTester.create_and_check_for_causal_lm with Llama->Dbrx
+    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 = DbrxForCausalLM(config=config)
+        model.to(torch_device)
+        model.eval()
+        result = model(input_ids, attention_mask=input_mask, labels=token_labels)
+        self.parent.assertEqual(result.logits.shape, (self.batch_size, self.seq_length, self.vocab_size))
+
+    def create_and_check_decoder_model_past_large_inputs(
+        self,
+        config,
+        input_ids,
+        token_type_ids,
+        input_mask,
+        sequence_labels,
+        token_labels,
+        choice_labels,
+        encoder_hidden_states,
+        encoder_attention_mask,
+    ):
+        config.is_decoder = True
+        config.add_cross_attention = True
+        model = DbrxForCausalLM(config=config)
+        model.to(torch_device)
+        model.eval()
+
+        # first forward pass
+        outputs = model(
+            input_ids,
+            attention_mask=input_mask,
+            encoder_hidden_states=encoder_hidden_states,
+            encoder_attention_mask=encoder_attention_mask,
+            use_cache=True,
+        )
+        past_key_values = outputs.past_key_values
+
+        # create hypothetical multiple next token and extent to next_input_ids
+        next_tokens = ids_tensor((self.batch_size, 3), config.vocab_size)
+        next_mask = ids_tensor((self.batch_size, 3), vocab_size=2)
+
+        # append to next input_ids and
+        next_input_ids = torch.cat([input_ids, next_tokens], dim=-1)
+        next_attention_mask = torch.cat([input_mask, next_mask], dim=-1)
+
+        output_from_no_past = model(
+            next_input_ids,
+            attention_mask=next_attention_mask,
+            encoder_hidden_states=encoder_hidden_states,
+            encoder_attention_mask=encoder_attention_mask,
+            output_hidden_states=True,
+        )["hidden_states"][0]
+        output_from_past = model(
+            next_tokens,
+            attention_mask=next_attention_mask,
+            encoder_hidden_states=encoder_hidden_states,
+            encoder_attention_mask=encoder_attention_mask,
+            past_key_values=past_key_values,
+            output_hidden_states=True,
+        )["hidden_states"][0]
+
+        # select random slice
+        random_slice_idx = ids_tensor((1,), output_from_past.shape[-1]).item()
+        output_from_no_past_slice = output_from_no_past[:, -3:, random_slice_idx].detach()
+        output_from_past_slice = output_from_past[:, :, random_slice_idx].detach()
+
+        self.parent.assertTrue(output_from_past_slice.shape[1] == next_tokens.shape[1])
+
+        # test that outputs are equal for slice
+        self.parent.assertTrue(torch.allclose(output_from_past_slice, output_from_no_past_slice, atol=1e-3))
+
+    # Copied from tests.models.llama.test_modeling_llama.LlamaModelTester.prepare_config_and_inputs_for_common with Llama->Dbrx
+    def prepare_config_and_inputs_for_common(self):
+        config_and_inputs = self.prepare_config_and_inputs()
+        (
+            config,
+            input_ids,
+            token_type_ids,
+            input_mask,
+            sequence_labels,
+            token_labels,
+            choice_labels,
+        ) = config_and_inputs
+        inputs_dict = {"input_ids": input_ids, "attention_mask": input_mask}
+        return config, inputs_dict
+
+
+@require_torch
+class DbrxModelTest(ModelTesterMixin, GenerationTesterMixin, PipelineTesterMixin, unittest.TestCase):
+    all_model_classes = (DbrxModel, DbrxForCausalLM) if is_torch_available() else ()
+    all_generative_model_classes = (DbrxForCausalLM,) if is_torch_available() else ()
+    pipeline_model_mapping = {"text-generation": DbrxForCausalLM} if is_torch_available() else {}
+    test_headmasking = False
+    test_pruning = False
+
+    def setUp(self):
+        self.model_tester = DbrxModelTester(self)
+        self.config_tester = ConfigTester(self, config_class=DbrxConfig, d_model=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)
+
+    @slow
+    def test_model_from_pretrained(self):
+        model_name = "eitanturok/dbrx-tiny"
+        model = DbrxModel.from_pretrained(model_name)
+        self.assertIsNotNone(model)
+
+    @unittest.skip("Dbrx models have weight tying disabled.")
+    def test_tied_weights_keys(self):
+        pass
+
+    @unittest.skip("TODO @gante fix this for Llama")
+    @parameterized.expand([(1, False), (1, True), (4, False)])
+    def test_new_cache_format(self, num_beams, do_sample):
+        pass
+
+
+@require_torch
+class DbrxModelIntegrationTest(unittest.TestCase):
+    @slow
+    def test_tiny_model_logits(self):
+        model = DbrxForCausalLM.from_pretrained("Rocketknight1/dbrx-tiny-random")
+        input_ids = torch.tensor([[0, 1, 2, 3, 4, 5]])
+        output = model(input_ids)[0]
+        vocab_size = model.vocab_size
+
+        expected_shape = torch.Size((1, 6, vocab_size))
+        self.assertEqual(output.shape, expected_shape)
+
+        expected_slice = torch.tensor(
+            [
+                [
+                    [-1.6300e-04, 5.0118e-04, 2.5437e-04],
+                    [2.0422e-05, 2.7210e-04, -1.5125e-04],
+                    [-1.5105e-04, 4.6879e-04, 3.3309e-04],
+                ]
+            ]
+        )
+        self.assertTrue(torch.allclose(output[:, :3, :3], expected_slice, atol=1e-4))