f90ce0b8fdbca174da1b904d6f72ba4c2624cd89bbe3461d64e899f5b7f65f03

LICENSE.txt

@@ -0,0 +1,176 @@
+Databricks Open Model License
+
+By using, reproducing, modifying, distributing, performing or displaying
+any portion or element of DBRX or DBRX Derivatives, or otherwise accepting
+the terms of this Agreement, you agree to be bound by this Agreement.
+
+Version Release Date: March 27, 2024
+
+
+Section 1: Definitions
+
+“Agreement” means these terms and conditions that govern the use, reproduction,
+modification, distribution, performance or display of DBRX and/or DBRX
+Derivatives and any terms and conditions incorporated by reference.
+
+“Databricks” or “we” means Databricks, Inc.
+
+“Licensee” or “you” means you, or your employer or any other person or entity
+(if you are entering into this Agreement on such person or entity’s behalf),
+of the age required under applicable laws, rules or regulations to provide
+legal consent and that has legal authority to bind your employer or such other
+person or entity if you are entering in this Agreement on their behalf.
+
+“DBRX Derivatives” means all (i) modifications to DBRX, (ii) works based on
+DBRX and (iii) any other derivative works thereof. Outputs are not deemed DBRX
+Derivatives.
+
+“DBRX” means the foundational large language models and software and
+algorithms, including machine-learning model code, trained model weights,
+inference-enabling code, training-enabling code, fine-tuning enabling code,
+documentation and other elements of the foregoing identified by Databricks at
+https://github.com/databricks/dbrx, regardless of the source that you obtained
+it from.
+
+“Output” means the results of operating DBRX or DBRX Derivatives.
+
+As used in this Agreement, “including” means “including without limitation.”
+
+
+Section 2: License Rights and Conditions on Use and Distribution
+
+2.1 Grant of Rights
+
+You are granted a non-exclusive, worldwide, non-transferable and royalty-free
+limited license under Databricks’ intellectual property or other rights owned
+by Databricks embodied in DBRX to use, reproduce, distribute, copy, modify,
+and create derivative works of DBRX in accordance with the terms of this
+Agreement.
+
+2.2 Reproduction and Distribution
+
+	1. All distributions of DBRX or DBRX Derivatives must be accompanied by a
+    "Notice" text file that contains the following notice: "DBRX is provided
+    under and subject to the Databricks Open Model License, Copyright ©
+    Databricks, Inc. All rights reserved."
+
+	2. If you distribute or make DBRX or DBRX Derivatives available to a third
+    party, you must provide a copy of this Agreement to such third party.
+
+	3. You must cause any modified files that you distribute to carry prominent
+    notices stating that you modified the files.
+
+You may add your own intellectual property statement to your modifications of
+DBRX and, except as set forth in this Section, may provide additional or
+different terms and conditions for use, reproduction, or distribution of DBRX
+or DBRX Derivatives as a whole, provided your use, reproduction, modification,
+distribution, performance, and display of DBRX or DBRX Derivatives otherwise
+complies with the terms and conditions of this Agreement. Any additional or
+different terms and conditions you impose must not conflict with the terms of
+this Agreement and in the event of a conflict, the terms and conditions of this
+Agreement shall govern over any such additional or different terms and conditions.
+
+2.3 Use Restrictions
+
+You will not use DBRX or DBRX Derivatives or any Output to improve any other
+large language model (excluding DBRX or DBRX Derivatives).
+
+You will not use DBRX or DBRX Derivatives:
+
+	1. for any restricted use set forth in the Databricks Open Model Acceptable
+    Use Policy identified at
+    https://www.databricks.com/legal/acceptable-use-policy-open-model
+    ("Acceptable Use Policy"), which is hereby incorporated by reference into
+    this Agreement; or
+
+	2. in violation of applicable laws and regulations.
+
+To the maximum extent permitted by law, Databricks reserves the right to
+restrict (remotely or otherwise) usage of DBRX or DBRX Derivatives that
+Databricks reasonably believes are in violation of this Agreement.
+
+
+Section 3: Additional Commercial Terms
+
+If, on the DBRX version release date, the monthly active users of the products
+or services made available by or for Licensee, or Licensee’s affiliates, is
+greater than 700 million monthly active users in the preceding calendar month,
+you must request a license from Databricks, which we may grant to you in our
+sole discretion, and you are not authorized to exercise any of the rights under
+this Agreement unless or until Databricks otherwise expressly grants you such
+rights.
+
+If you receive DBRX or DBRX Derivatives from a direct or indirect licensee as
+part of an integrated end user product, then this section (Section 3) of the
+Agreement will not apply to you.
+
+
+Section 4: Additional Provisions
+
+4.1 Updates
+
+Databricks may update DBRX from time to time, and you must make reasonable
+efforts to use the latest version of DBRX.
+
+4.2 Intellectual Property
+
+a. No trademark licenses are granted under this Agreement, and in connection
+with DBRX or DBRX Derivatives, neither Databricks nor Licensee may use any name
+or mark owned by or associated with the other or any of its affiliates, except
+as required for reasonable and customary use in describing and redistributing
+DBRX or DBRX Derivatives.
+
+b. Subject to Databricks’ ownership of DBRX and DRBX Derivatives made by or for
+Databricks, with respect to any DBRX Derivatives that are made by you, as
+between you and Databricks, you are and will be the owner of such DBRX
+Derivatives.
+
+c. Databricks claims no ownership rights in Outputs. You are responsible for
+Outputs and their subsequent uses.
+
+d. If you institute litigation or other proceedings against Databricks or any
+entity (including a cross-claim or counterclaim in a lawsuit) alleging that
+DBRX or Outputs or results therefrom, or any portion of any of the foregoing,
+constitutes infringement of intellectual property or other rights owned or
+licensable by you, then any licenses granted to you under this Agreement shall
+terminate as of the date such litigation or claim is filed or instituted. You
+will indemnify and hold harmless Databricks from and against any claim by any
+third party arising out of or related to your use or distribution of DBRX or
+DBRX Derivatives.
+
+4.3 DISCLAIMER OF WARRANTY
+
+UNLESS REQUIRED BY APPLICABLE LAW, DBRX AND ANY OUTPUT AND RESULTS THEREFROM
+ARE PROVIDED ON AN “AS IS” BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER
+EXPRESS OR IMPLIED, INCLUDING, WITHOUT LIMITATION, ANY WARRANTIES OF TITLE,
+NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE. YOU
+ARE SOLELY RESPONSIBLE FOR DETERMINING THE APPROPRIATENESS OF USING OR
+REDISTRIBUTING DBRX OR DBRX DERIVATIVES AND ANY OUTPUT AND ASSUME ANY RISKS
+ASSOCIATED WITH YOUR USE OF DBRX OR DBRX DERIVATIVES AND ANY OUTPUT AND RESULTS.
+
+4.4 LIMITATION OF LIABILITY
+
+IN NO EVENT WILL DATABRICKS OR ITS AFFILIATES BE LIABLE UNDER ANY THEORY OF
+LIABILITY, WHETHER IN CONTRACT, TORT, NEGLIGENCE, PRODUCTS LIABILITY, OR
+OTHERWISE, ARISING OUT OF THIS AGREEMENT, FOR ANY LOST PROFITS OR ANY INDIRECT,
+SPECIAL, CONSEQUENTIAL, INCIDENTAL, EXEMPLARY OR PUNITIVE DAMAGES, EVEN IF
+DATABRICKS OR ITS AFFILIATES HAVE BEEN ADVISED OF THE POSSIBILITY OF ANY OF THE
+FOREGOING.
+
+4.5 Term and Termination
+
+The term of this Agreement will commence upon your acceptance of this Agreement
+or access to DBRX or DBRX Derivatives and will continue in full force and
+effect until terminated in accordance with the terms and conditions herein.
+Databricks may terminate this Agreement if you are in breach of any term or
+condition of this Agreement. Upon termination of this Agreement, you shall
+delete and cease use of DBRX or any DBRX Derivatives. Sections 1, 4.2(d), 4.3,
+4.4, and 4.6 shall survive the termination of this Agreement.
+
+4.6 Governing Law and Jurisdiction
+
+This Agreement will be governed and construed under the laws of the State of
+California without regard to choice of law principles, and the UN Convention
+on Contracts for the International Sale of Goods does not apply to this
+Agreement. The courts of California shall have exclusive jurisdiction of any
+dispute arising out of this Agreement.

NOTICE.txt

@@ -0,0 +1 @@
+DBRX is provided under and subject to the Databricks Open Model License, Copyright © Databricks, Inc. All rights reserved.

README.md

@@ -1,3 +1,172 @@
----
-license: unknown
----
+---
+extra_gated_heading: You need to share contact information with Databricks to access this model
+extra_gated_prompt: >-
+
+  ### DBRX Terms of Use
+
+  Use of DBRX is governed by the [Databricks Open Model License](https://www.databricks.com/legal/open-model-license) and the [Databricks Open Model Acceptable Use Policy](https://www.databricks.com/legal/acceptable-use-policy-open-model).
+  
+extra_gated_fields:
+  First Name: text
+  Last Name: text
+  Organization: text
+  Purpose for Base Model Access: text
+  By clicking 'Submit' below, I accept the terms of the license and acknowledge that the information I provide will be collected, stored, processed, and shared in accordance with Databricks' Privacy Notice and I understand I can update my preferences at any time: checkbox
+extra_gated_description: >-
+  The information you provide will be collected, stored, processed, and shared in accordance with Databricks [Privacy Notice](https://www.databricks.com/legal/privacynotice).
+extra_gated_button_content: Submit
+inference: false
+license: other
+license_name: databricks-open-model-license
+license_link: https://www.databricks.com/legal/open-model-license
+---
+
+# DBRX Base
+
+* DBRX Base is a mixture-of-experts (MoE) large language model trained from scratch by Databricks.
+* We are releasing both DBRX Base, a pretrained base model, and DBRX Instruct, a fine-tuned version for few-turn interactions, under [an open license](https://www.databricks.com/legal/open-model-license).
+* This is the repository for DBRX Base. DBRX Instruct can be found [here](https://huggingface.co/databricks/dbrx-instruct).
+* For full details on the DBRX models, please read our [technical blog post](https://www.databricks.com/blog/introducing-dbrx-new-state-art-open-llm).
+
+
+## Model Overview
+DBRX is a [transformer-based](https://www.isattentionallyouneed.com/) decoder-only large language model (LLM) that was trained using next-token prediction. 
+It uses a *fine-grained* mixture-of-experts (MoE) architecture with 132B total parameters of which 36B parameters are active on any input. 
+It was pre-trained on 12T tokens of text and code data. 
+Compared to other open MoE models like Mixtral-8x7B and Grok-1, DBRX is fine-grained, meaning it uses a larger number of smaller experts. DBRX has 16 experts and chooses 4, while Mixtral-8x7B and Grok-1 have 8 experts and choose 2. 
+This provides 65x more possible combinations of experts and we found that this improves model quality. 
+DBRX uses rotary position encodings (RoPE), gated linear units (GLU), and grouped query attention (GQA). 
+It uses the GPT-4 tokenizer as provided in the [tiktoken](https://github.com/openai/tiktoken) repository. 
+We made these choices based on exhaustive evaluation and scaling experiments.
+
+DBRX was pretrained on 12T tokens of carefully curated data and a maximum context length of 32K tokens. 
+We estimate that this data is at least 2x better token-for-token than the data we used to pretrain the MPT family of models. 
+This new dataset was developed using the full suite of Databricks tools, including Apache Spark™ and Databricks notebooks for data processing, and Unity Catalog for data management and governance. 
+We used curriculum learning for pretraining, changing the data mix during training in ways we found to substantially improve model quality.
+
+* **Inputs:** DBRX only accepts text-based inputs and accepts a context length of up to 32768 tokens.
+* **Outputs:** DBRX only produces text-based outputs.  
+* **Model Architecture:** More detailed information about DBRX Instruct and DBRX Base can be found in our [technical blog post](https://www.databricks.com/blog/introducing-dbrx-new-state-art-open-llm).
+* **License:** [Databricks Open Model License](https://www.databricks.com/legal/open-model-license)
+* **Acceptable Use Policy:** [Databricks Open Model Acceptable Use Policy](https://www.databricks.com/legal/acceptable-use-policy-open-model)
+* **Version:** 1.0
+* **Owner:** Databricks, Inc.
+
+
+## Usage
+These are several general ways to use the DBRX models: 
+* DBRX Base and DBRX Instruct are available for download on HuggingFace (see our Quickstart guide below). This is the HF repository for DBRX Base; DBRX Instruct can be found [here](https://huggingface.co/databricks/dbrx-instruct). 
+* The DBRX model repository can be found on GitHub [here](https://github.com/databricks/dbrx). 
+* DBRX Base and DBRX Instruct are available with [Databricks Foundation Model APIs](https://docs.databricks.com/en/machine-learning/foundation-models/index.html) via both *Pay-per-token* and *Provisioned Throughput* endpoints. These are enterprise-ready deployments.
+* For more information on how to fine-tune using LLM-Foundry, please take a look at our LLM pretraining and fine-tuning [documentation](https://github.com/mosaicml/llm-foundry/blob/main/scripts/train/README.md).
+
+
+## Quickstart Guide
+**NOTE: This is DBRX Base, and has not been instruction finetuned. It has not been trained for interactive chat and is only a completion model.**
+If you are looking for the finetuned model, please use [DBRX Instruct](https://huggingface.co/databricks/dbrx-instruct).
+
+Getting started with DBRX models is easy with the `transformers` library. The model requires ~264GB of RAM and the following packages:
+
+```bash
+pip install "transformers>=4.39.2" "tiktoken>=0.6.0"
+```
+
+If you'd like to speed up download time, you can use the `hf_transfer` package as described by Huggingface [here](https://huggingface.co/docs/huggingface_hub/en/guides/download#faster-downloads).
+```bash
+pip install hf_transfer
+export HF_HUB_ENABLE_HF_TRANSFER=1
+```
+
+You will need to request access to this repository to download the model. Once this is granted, 
+[obtain an access token](https://huggingface.co/docs/hub/en/security-tokens) with `read` permission, and supply the token below.
+
+### Run the model on a CPU:
+```python
+from transformers import AutoTokenizer, AutoModelForCausalLM
+import torch
+
+tokenizer = AutoTokenizer.from_pretrained("databricks/dbrx-base", trust_remote_code=True, token="hf_YOUR_TOKEN")
+model = AutoModelForCausalLM.from_pretrained("databricks/dbrx-base", device_map="cpu", torch_dtype=torch.bfloat16, trust_remote_code=True, token="hf_YOUR_TOKEN")
+
+input_text = "Databricks was founded in "
+input_ids = tokenizer(input_text, return_tensors="pt")
+
+outputs = model.generate(**input_ids, max_new_tokens=100)
+print(tokenizer.decode(outputs[0]))
+```
+
+### Run the model on multiple GPUs:
+```python
+from transformers import AutoTokenizer, AutoModelForCausalLM
+import torch
+
+tokenizer = AutoTokenizer.from_pretrained("databricks/dbrx-base", trust_remote_code=True, token="hf_YOUR_TOKEN")
+model = AutoModelForCausalLM.from_pretrained("databricks/dbrx-base", device_map="auto", torch_dtype=torch.bfloat16, trust_remote_code=True, token="hf_YOUR_TOKEN")
+
+input_text = "Databricks was founded in "
+input_ids = tokenizer(input_text, return_tensors="pt").to("cuda")
+
+outputs = model.generate(**input_ids, max_new_tokens=100)
+print(tokenizer.decode(outputs[0]))
+```
+If your GPU system supports [FlashAttention2](https://huggingface.co/docs/transformers/perf_infer_gpu_one#flashattention-2), you can add `attn_implementation=”flash_attention_2”` as a keyword to `AutoModelForCausalLM.from_pretrained()` to achieve faster inference.
+
+
+## Limitations and Ethical Considerations
+### Training Dataset Limitations
+The DBRX models were trained on 12T tokens of text, with a knowledge cutoff date of December 2023.
+
+The training mix used for DBRX contains both natural-language and code examples. The vast majority of our training data is in the English language. We did not test DBRX for non-English proficiency. Therefore, DBRX should be considered a generalist model for text-based use in the English language.
+
+DBRX does not have multimodal capabilities.
+
+### Associated Risks and Recommendations 
+All foundation models are novel technologies that carry various risks, and may output information that is inaccurate, incomplete, biased, or offensive. 
+Users should exercise judgment and evaluate such output for accuracy and appropriateness for their desired use case before using or sharing it. 
+Databricks recommends [using retrieval augmented generation (RAG)](https://www.databricks.com/glossary/retrieval-augmented-generation-rag) in scenarios where accuracy and fidelity are important. 
+We also recommend that anyone using or fine-tuning either DBRX Base or DBRX Instruct perform additional testing around safety in the context of their particular application and domain. 
+
+
+## Intended Uses
+### Intended Use Cases
+The DBRX models are open, general-purpose LLMs intended and licensed for both commercial and research applications. 
+They can be further fine-tuned for various domain-specific natural language and coding tasks. 
+DBRX Base can be used as an off-the-shelf model for text completion for general English-language and coding tasks. 
+
+Please review the Associated Risks section above, as well as the [Databricks Open Model License](https://www.databricks.com/legal/open-model-license) and [Databricks Open Model Acceptable Use Policy](https://www.databricks.com/legal/acceptable-use-policy-open-model) for further information about permissible uses of DBRX Base and its derivatives. 
+
+### Out-of-Scope Use Cases
+DBRX models are not intended to be used out-of-the-box in non-English languages and do not support native code execution, or other forms of function-calling. 
+DBRX models should not be used in any manner that violates applicable laws or regulations or in any other way that is prohibited by the [Databricks Open Model License](https://www.databricks.com/legal/open-model-license) and [Databricks Open Model Acceptable Use Policy](https://www.databricks.com/legal/acceptable-use-policy-open-model). 
+
+
+## Training Stack
+MoE models are complicated to train, and the training of DBRX Base and DBRX Instruct was heavily supported by Databricks’ infrastructure for data processing and large-scale LLM training (e.g., [Composer](https://github.com/mosaicml/composer), [Streaming](https://github.com/mosaicml/streaming), [Megablocks](https://github.com/stanford-futuredata/megablocks), and [LLM Foundry](https://github.com/mosaicml/llm-foundry)). 
+
+Composer is our core library for large-scale training. 
+It provides an optimized training loop, easy [checkpointing](https://docs.mosaicml.com/projects/composer/en/latest/trainer/checkpointing.html) and [logging](https://docs.mosaicml.com/projects/composer/en/latest/trainer/logging.html#wood-logging), 
+[FSDP](https://pytorch.org/docs/stable/fsdp.html)-based [model sharding](https://docs.mosaicml.com/projects/composer/en/latest/notes/distributed_training.html#fullyshardeddataparallel-fsdp), 
+convenient [abstractions](https://docs.mosaicml.com/projects/composer/en/latest/trainer/time.html), extreme customizability via [callbacks](https://docs.mosaicml.com/projects/composer/en/latest/trainer/callbacks.html), and more.
+
+Streaming enables fast, low cost, and scalable training on large datasets from cloud storage. It handles a variety of challenges around deterministic resumption as node counts change, avoiding redundant downloads across devices, high-quality shuffling at scale, sample-level random access, and speed.
+
+Megablocks is a lightweight library for MoE training. Crucially, it supports “dropless MoE,” which avoids inefficient padding and is intended to provide deterministic outputs for a given sequence no matter what other sequences are in the batch.
+
+LLM Foundry ties all of these libraries together to create a simple LLM pretraining, fine-tuning, and inference experience.
+
+DBRX was trained using proprietary optimized versions of the above open source libraries, along with our [LLM training platform](https://www.databricks.com/product/machine-learning/mosaic-ai-training). 
+
+
+## Evaluation
+We find that DBRX outperforms established open-source and open-weight base models on the [Databricks Model Gauntlet](https://www.databricks.com/blog/llm-evaluation-for-icl), the [Hugging Face Open LLM Leaderboard](https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard), and HumanEval. 
+The Databricks Model Gauntlet measures performance on more than 30 tasks across six categories: world knowledge, common sense reasoning, language understanding, reading comprehension, symbolic problem solving, and programming. 
+The Hugging Face Open LLM Leaderboard measures the average of ARC-Challenge, HellaSwag, MMLU, TruthfulQA, Winogrande and GSM8k. 
+HumanEval measures coding ability.
+
+Full evaluation details can be found in our [technical blog post](https://www.databricks.com/blog/introducing-dbrx-new-state-art-open-llm). 
+
+
+## Acknowledgements
+The DBRX models were made possible thanks in large part to the open-source community, especially:
+* The [MegaBlocks](https://arxiv.org/abs/2211.15841) library, which established a foundation for our MoE implementation.
+* [PyTorch FSDP](https://arxiv.org/abs/2304.11277), which we built on for distributed training.

config.json

@@ -0,0 +1,38 @@
+{
+  "architectures": [
+    "DbrxForCausalLM"
+  ],
+  "attn_config": {
+    "clip_qkv": 8,
+    "kv_n_heads": 8,
+    "model_type": "",
+    "rope_theta": 500000
+  },
+  "auto_map": {
+    "AutoConfig": "configuration_dbrx.DbrxConfig",
+    "AutoModelForCausalLM": "modeling_dbrx.DbrxForCausalLM"
+  },
+  "d_model": 6144,
+  "emb_pdrop": 0.0,
+  "ffn_config": {
+    "ffn_hidden_size": 10752,
+    "model_type": "",
+    "moe_jitter_eps": 0.01,
+    "moe_loss_weight": 0.05,
+    "moe_num_experts": 16,
+    "moe_top_k": 4
+  },
+  "initializer_range": 0.02,
+  "max_seq_len": 32768,
+  "model_type": "dbrx_converted",
+  "n_heads": 48,
+  "n_layers": 40,
+  "output_router_logits": false,
+  "resid_pdrop": 0.0,
+  "router_aux_loss_coef": 0.05,
+  "tie_word_embeddings": false,
+  "torch_dtype": "bfloat16",
+  "transformers_version": "4.38.2",
+  "use_cache": true,
+  "vocab_size": 100352
+}

configuration_dbrx.py

@@ -0,0 +1,264 @@
+"""Dbrx configuration."""
+from typing import Any, Optional
+
+from transformers.configuration_utils import PretrainedConfig
+from transformers.utils import logging
+
+logger = logging.get_logger(__name__)
+
+DBRX_PRETRAINED_CONFIG_ARCHIVE_MAP = {}
+
+
+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*, defualts to None):
+            If not `None`, clip the queries, keys, and values in the attention layer to this value.
+        kv_n_heads (Optional[int]): For grouped_query_attention only, allow user to specify number of kv heads.
+        rope_theta (float): The base frequency for rope.
+    """
+
+    def __init__(
+        self,
+        attn_pdrop: float = 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): 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.
+        ffn_hidden_size (int, optional): The hidden size of the feedforward network.
+        moe_num_experts (int, optional): The number of experts in the mixture of experts layer.
+        moe_top_k (int, optional): The number of experts to use in the mixture of experts layer.
+        moe_jitter_eps (float, optional): The jitter epsilon for the mixture of experts layer.
+        moe_loss_weight (float, optional): The loss weight for the mixture of experts layer.
+        moe_normalize_expert_weights (float, optional): The normalization factor for the expert weights.
+        uniform_expert_assignment (bool, optional): Whether to use uniform expert assignment.
+            This should only be used for benchmarking purposes.
+    """
+
+    def __init__(
+        self,
+        ffn_act_fn: Optional[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,
+        uniform_expert_assignment: bool = False,
+        **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
+        self.uniform_expert_assignment = uniform_expert_assignment
+
+        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):
+    """Configuration class for Dbrx.
+
+    [`DbrxModel`]. It is used to instantiate a Dbrx model according to the
+    specified arguments, defining the model 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 6144):
+            Dimensionality of the embeddings and hidden states.
+        n_heads (`int`, *optional*, defaults to 48):
+            Number of attention heads for each attention layer in the Transformer encoder.
+        n_layers (`int`, *optional*, defaults to 40):
+            Number of hidden layers in the Transformer encoder.
+        max_seq_len (`int`, *optional*, defaults to 32768):
+            The maximum sequence length of the model.
+        vocab_size (`int`, *optional*, defaults to 100352):
+            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 `False`):
+            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
+        router_aux_loss_coef (`float`, *optional*, defaults to 0.001):
+            The aux loss factor for the total loss.
+
+
+    Example:
+    ```python
+    >>> from transformers import DbrxConfig, DbrxModel
+
+    >>> # Initializing a Dbrx configuration
+    >>> configuration = DbrxConfig()
+
+    >>> # 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,
+        router_aux_loss_coef: float = 0.05,
+        **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
+        self.router_aux_loss_coef = router_aux_loss_coef
+
+        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,
+        )

generation_config.json

@@ -0,0 +1,7 @@
+{
+  "_from_model_config": true,
+  "eos_token_id": [
+    100257
+  ],
+  "transformers_version": "4.38.2"
+}

modeling_dbrx.py

@@ -0,0 +1,1479 @@
+"""PyTorch Dbrx model."""
+
+import math
+import warnings
+from copy import deepcopy
+from functools import partial
+from typing import Any, Callable, Dict, Optional, Tuple, Union
+
+import torch
+import torch.nn.functional as F
+import torch.utils.checkpoint
+from torch import nn
+from transformers.cache_utils import Cache, DynamicCache, StaticCache
+from transformers.modeling_attn_mask_utils import AttentionMaskConverter
+from transformers.modeling_outputs import (MoeCausalLMOutputWithPast,
+                                           MoeModelOutputWithPast)
+from transformers.modeling_utils import PreTrainedModel
+from transformers.utils import is_flash_attn_2_available, logging
+
+from .configuration_dbrx import DbrxAttentionConfig, DbrxConfig, DbrxFFNConfig
+
+if is_flash_attn_2_available():
+    try:
+        from flash_attn import flash_attn_func, flash_attn_varlen_func
+        from flash_attn.bert_padding import pad_input  # noqa
+        from flash_attn.bert_padding import index_first_axis, unpad_input
+    except:
+        pass
+
+logger = logging.get_logger(__name__)
+
+_CONFIG_FOR_DOC = 'DbrxConfig'
+
+#############################################################################
+# Copied from LLaMaRotaryEmbedding
+#############################################################################
+
+
+class DbrxRotaryEmbedding(nn.Module):
+
+    def __init__(self,
+                 dim: int,
+                 max_position_embeddings: int = 2048,
+                 base: float = 10000.0,
+                 scaling_factor: float = 1.0):
+        super().__init__()
+        self.scaling_factor = scaling_factor
+        self.dim = dim
+        self.max_position_embeddings = max_position_embeddings
+        self.base = base
+        inv_freq = 1.0 / (self.base**(
+            torch.arange(0, self.dim, 2, dtype=torch.int64).float() / self.dim))
+        self.register_buffer('inv_freq', inv_freq, persistent=False)
+        # For BC we register cos and sin cached
+        self.max_seq_len_cached = max_position_embeddings
+
+    @torch.no_grad()
+    def forward(
+            self, x: torch.Tensor, position_ids: torch.LongTensor
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        # x: [bs, num_attention_heads, seq_len, head_size]
+        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)
+
+
+def rotate_half(x: torch.Tensor) -> torch.Tensor:
+    """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)
+
+
+def apply_rotary_pos_emb(
+        q: torch.Tensor,
+        k: torch.Tensor,
+        cos: torch.Tensor,
+        sin: torch.Tensor,
+        unsqueeze_dim: int = 1) -> Tuple[torch.Tensor, torch.Tensor]:
+    """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.
+        unsqueeze_dim (`int`, *optional*, defaults to 1):
+            The 'unsqueeze_dim' argument specifies the dimension along which to unsqueeze cos and
+            sin so that they can be properly broadcasted to the dimensions of q and k. For example, note
+            that cos and sin 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 and sin 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
+
+
+def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
+    """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)
+
+
+#############################################################################
+
+#############################################################################
+# Modified from modeling_mixtral
+#############################################################################
+
+
+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
+
+
+#############################################################################
+
+
+def resolve_ffn_act_fn(
+        ffn_act_fn: dict) -> Callable[[torch.Tensor], torch.Tensor]:
+    """Resolve the activation function for the feed-forward network.
+
+    Args:
+        ffn_act_fn (dict): The configuration dictionary for the activation function.
+            The dict config must specify the 'name' of a torch.nn.functional activation
+            function. All of other key values pairs are bound to the function as a partial.
+
+    Returns:
+        Callable[[torch.Tensor], torch.Tensor]: The activation function.
+    """
+    config = deepcopy(ffn_act_fn)
+    name = config.pop('name')
+    if not hasattr(nn.functional, name):
+        raise ValueError(f'Unrecognised activation function name ({name}).')
+    act = getattr(nn.functional, name)
+    return partial(act, **config)
+
+
+#############################################################################
+# Copied from LLaMaAttention
+#############################################################################
+
+def get_max_seqlen_in_batch(attention_mask):
+    max_num = torch.max(attention_mask)
+    # attention_mask: B x N
+    counts = []
+    for i in range(1, max_num + 1):
+        counts.append(
+            torch.sum(attention_mask == i, axis=-1)
+        )  # shape: B, count length of data point maksed with i
+    result = torch.stack(counts, axis=1)
+    result = result.flatten()
+    return result[result.nonzero()].squeeze(-1).to(dtype=torch.int32)
+
+
+def _get_unpad_data(attention_mask):
+    seqlens_in_batch = get_max_seqlen_in_batch(
+        attention_mask
+    )  # 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.torch.int32), (1, 0)
+    )
+    return (
+        indices,
+        cu_seqlens,
+        max_seqlen_in_batch,
+    )
+
+
+class DbrxAttention(nn.Module):
+    """Multi-head self attention."""
+
+    def __init__(self,
+                 hidden_size: int,
+                 num_heads: int,
+                 max_position_embeddings: int,
+                 attn_config: DbrxAttentionConfig,
+                 block_idx: Optional[int] = None):
+        super().__init__()
+        self.hidden_size = hidden_size
+        self.num_heads = num_heads
+        self.head_dim = self.hidden_size // self.num_heads
+        self.max_position_embeddings = max_position_embeddings
+        self.block_idx = block_idx
+        self.config = attn_config
+        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.')
+
+        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.q_proj = nn.Linear(self.hidden_size,
+                              self.hidden_size,
+                              bias=False)
+        self.k_proj = nn.Linear(self.hidden_size,
+                              self.num_key_value_heads * self.head_dim,
+                              bias=False)
+        self.v_proj = nn.Linear(self.hidden_size,
+                              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()
+
+        query_states = self.q_proj(hidden_states)
+        key_states = self.k_proj(hidden_states)
+        value_states = self.v_proj(hidden_states)
+        if self.clip_qkv is not None:
+            query_states = query_states.clamp(min=-self.clip_qkv, max=self.clip_qkv)
+            key_states = key_states.clamp(min=-self.clip_qkv, max=self.clip_qkv)
+            value_states = value_states.clamp(min=-self.clip_qkv, max=self.clip_qkv)
+
+        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):
+        if not is_flash_attn_2_available():
+            raise ImportError(
+                'Flash Attention 2 is not available. Please install it with `pip install flash-attn`.'
+            )
+
+        super().__init__(*args, **kwargs)
+
+    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()
+
+        query_states = self.q_proj(hidden_states)
+        key_states = self.k_proj(hidden_states)
+        value_states = self.v_proj(hidden_states)
+        if self.clip_qkv is not None:
+            query_states = query_states.clamp(min=-self.clip_qkv, max=self.clip_qkv)
+            key_states = key_states.clamp(min=-self.clip_qkv, max=self.clip_qkv)
+            value_states = value_states.clamp(min=-self.clip_qkv, max=self.clip_qkv)
+
+        # 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(
+                f'The input hidden states seems to be silently casted in float32, this might be '
+                +
+                f'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  # type: ignore
+
+    def _flash_attention_forward(
+        self,
+        query_states: torch.Tensor,
+        key_states: torch.Tensor,
+        value_states: torch.Tensor,
+        attention_mask: Union[torch.LongTensor, None],
+        query_length: int,
+        dropout: float = 0.0,
+        softmax_scale: Optional[float] = None,
+    ):
+        """Use FlashAttention, stripping padding tokens if necessary.
+
+        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.LongTensor | None): 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.
+            query_length (int): The length of the query sequence
+            dropout (float): Attention dropout
+            softmax_scale (float, optional): The scaling of QK^T before applying softmax.
+                Defaults to 1 / sqrt(head_dim)
+        """
+        causal = True
+        # 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
+
+    def _upad_input(self, query_layer: torch.Tensor, key_layer: torch.Tensor,
+                    value_layer: torch.Tensor, attention_mask: torch.Tensor,
+                    query_length: int):
+        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),
+        )
+
+
+DBRX_ATTENTION_CLASSES = {
+    'eager': DbrxAttention,
+    'flash_attention_2': DbrxFlashAttention2,
+}
+
+
+class DbrxNormAttentionNorm(nn.Module):
+
+    def __init__(
+        self,
+        hidden_size: int,
+        num_heads: int,
+        max_position_embeddings: int,
+        resid_pdrop: float,
+        attn_implementation: str,
+        attn_config: DbrxAttentionConfig,
+        block_idx: Optional[int] = None,
+    ):
+        super().__init__()
+        self.block_idx = block_idx
+        self.resid_pdrop = resid_pdrop
+        self.norm_1 = nn.LayerNorm(hidden_size, bias=False)
+        self.attn = DBRX_ATTENTION_CLASSES[attn_implementation](
+            hidden_size=hidden_size,
+            num_heads=num_heads,
+            max_position_embeddings=max_position_embeddings,
+            attn_config=attn_config,
+            block_idx=block_idx,
+        )
+        self.norm_2 = nn.LayerNorm(hidden_size, bias=False)
+
+        # print(f"Freezing norm_1/2 layers...")
+        self.norm_1.requires_grad_(False)
+        self.norm_2.requires_grad_(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
+
+        # with torch.no_grad():
+        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
+
+        # with torch.no_grad():
+        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],
+                 uniform_expert_assignment: bool):
+        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.uniform_expert_assignment = uniform_expert_assignment
+
+        self.layer = nn.Linear(self.hidden_size,
+                               self.moe_num_experts,
+                               bias=False)
+
+        # freeze router for training
+        # print("Freezing DbrxRouter Layer")
+        self.layer.requires_grad_(False)
+        # same as layer.requires_grad_(False)
+        # for param in self.parameters():
+        #     param.requires_grad_(False)
+        #     print(param)
+
+
+    def jitter(self, x: torch.Tensor) -> torch.Tensor:
+        if self.moe_jitter_eps is None:
+            raise RuntimeError('The router does not have moe_jitter_eps set.')
+        low = 1.0 - self.moe_jitter_eps
+        high = 1.0 + self.moe_jitter_eps
+        noise = torch.rand(x.size(), dtype=x.dtype, device=x.device)
+        return low + noise * (high - low)
+
+    def forward(
+            self, x: torch.Tensor
+    ) -> Tuple[torch.Tensor, torch.Tensor, torch.LongTensor]:
+        if self.training and self.moe_jitter_eps is not None:
+            x = x * self.jitter(x)
+
+        weights = self.layer(x.view(-1,
+                                    x.shape[-1])).softmax(dim=-1,
+                                                          dtype=torch.float32)
+        top_weights, top_experts = torch.topk(weights, self.moe_top_k, dim=-1)
+
+        if self.moe_normalize_expert_weights:
+            top_weights = top_weights / torch.norm(
+                top_weights,
+                p=self.moe_normalize_expert_weights,
+                dim=-1,
+                keepdim=True)
+
+        if self.uniform_expert_assignment:
+            with torch.no_grad():
+                uniform_tensor = torch.arange(
+                    0,
+                    top_experts.numel(),
+                    device=top_experts.device,
+                    dtype=top_experts.dtype) % self.moe_num_experts
+                top_experts = uniform_tensor.reshape(top_experts.shape)
+                # Note, weights and top_weights are not changed
+
+        weights = weights.to(x.dtype)
+        top_weights = top_weights.to(x.dtype)
+        return weights, top_weights, top_experts  # type: ignore
+
+
+class DbrxMLP(nn.Module):
+
+    def __init__(self, hidden_size: int, ffn_hidden_size: int, ffn_act_fn: dict):
+        super().__init__()
+
+        self.w1 = nn.Linear(hidden_size, ffn_hidden_size, bias=False)
+        self.v1 = nn.Linear(hidden_size, ffn_hidden_size, bias=False)
+        self.w2 = nn.Linear(ffn_hidden_size, hidden_size, bias=False)
+        self.activation_fn = resolve_ffn_act_fn(ffn_act_fn)
+    
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+
+        return self.w2(self.activation_fn(self.w1(x)) * self.v1(x))
+
+
+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 = nn.ModuleList([DbrxMLP(hidden_size, ffn_hidden_size, ffn_act_fn) for _ in range(moe_num_experts)])
+
+    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)
+        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
+
+            expert_tokens = x[None, token_idx].reshape(-1, hidden_size)
+            expert_out = self.mlp[expert_idx](expert_tokens) * top_weights[token_idx, topk_idx, 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, hidden_size: int, ffn_config: DbrxFFNConfig):
+        super().__init__()
+
+        self.router = DbrxRouter(
+            hidden_size,
+            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,
+            uniform_expert_assignment=ffn_config.uniform_expert_assignment,
+        )
+
+        self.experts = DbrxExperts(
+            hidden_size=hidden_size,
+            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]:
+        with torch.no_grad():
+            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(
+            hidden_size=config.d_model,
+            num_heads=config.n_heads,
+            max_position_embeddings=config.max_seq_len,
+            resid_pdrop=config.resid_pdrop,
+            attn_implementation=config._attn_implementation,
+            attn_config=config.attn_config,
+            block_idx=block_idx,
+        )
+        self.ffn = DbrxFFN(hidden_size=config.d_model,
+                           ffn_config=config.ffn_config)
+
+    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: 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
+        """
+        if 'padding_mask' in kwargs:
+            warnings.warn(
+                'Passing `padding_mask` is deprecated and will be removed in v4.37. Please make sure use `attention_mask` instead.`'
+            )
+
+        # 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
+
+
+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 = False
+    _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_()
+
+    def _setup_cache(self, cache_cls: Any, max_batch_size: int,
+                     max_cache_len: int):  # TODO: how to set var type of class?
+        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
+
+
+class DbrxModel(DbrxPreTrainedModel):
+    """Transformer decoder consisting of *config.num_hidden_layers*
+
+    [`DbrxBlock`] layers.
+
+    Args:
+        config: DbrxConfig
+    """
+
+    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
+
+    def _autocast_input_embeddings(self,
+                                   inputs_embeds: torch.Tensor) -> torch.Tensor:
+        if inputs_embeds.device.type == 'cuda' and torch.is_autocast_enabled():
+            return inputs_embeds.to(dtype=torch.get_autocast_gpu_dtype())
+        elif inputs_embeds.device.type == 'cpu' and torch.is_autocast_cpu_enabled(
+        ):
+            return inputs_embeds.to(dtype=torch.get_autocast_cpu_dtype())
+        else:
+            return inputs_embeds
+
+    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 = self._autocast_input_embeddings(
+            inputs_embeds)  # type: ignore
+        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(  # type: ignore
+                )
+
+        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(  # type: ignore
+                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)  # type: ignore
+
+        causal_mask = self._update_causal_mask(attention_mask, inputs_embeds,
+                                               cache_position)  # type: ignore
+
+        # 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,)  # type: ignore
+
+            if self.gradient_checkpointing and self.training:
+                block_outputs = self._gradient_checkpointing_func(
+                    block.__call__,
+                    hidden_states,
+                    attention_mask=causal_mask,
+                    position_ids=position_ids,
+                    past_key_values=past_key_values,
+                    output_attentions=output_attentions,
+                    output_router_logits=output_router_logits,
+                    use_cache=use_cache,
+                    cache_position=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],)  # type: ignore
+
+            if output_router_logits:
+                all_router_logits += (block_outputs[-1],)  # type: ignore
+
+        hidden_states = self.norm_f(hidden_states)
+
+        # add hidden states from the last decoder layer
+        if output_hidden_states:
+            all_hidden_states += (hidden_states,)  # type: ignore
+
+        next_cache = None
+        if use_cache:
+            next_cache = (
+                next_decoder_cache.to_legacy_cache()  # type: ignore
+                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  # type: ignore
+                (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
+
+
+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.router_aux_loss_coef = config.router_aux_loss_coef
+        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
+
+    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.
+
+        Example:
+        ```python
+        >>> from transformers import AutoTokenizer, DbrxForCausalLM
+
+        >>> model = DbrxForCausalLM.from_pretrained("databricks/dbrx")
+        >>> tokenizer = AutoTokenizer.from_pretrained("databricks/dbrx")
+
+        >>> 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.router_aux_loss_coef * aux_loss.to(
+                    loss.device)  # make sure to reside in the same device
+
+        if not return_dict:
+            output = (logits,) + outputs[1:]
+            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(
+            { # type: ignore
+                '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

special_tokens_map.json

@@ -0,0 +1,30 @@
+{
+  "bos_token": {
+    "content": "<|endoftext|>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "eos_token": {
+    "content": "<|endoftext|>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "pad_token": {
+    "content": "<|endoftext|>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "unk_token": {
+    "content": "<|endoftext|>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  }
+}

tiktoken.py

@@ -0,0 +1,374 @@
+"""Dbrx tokenizer."""
+
+from functools import lru_cache
+from typing import Any, Dict, List, Optional, Tuple
+
+from transformers import PreTrainedTokenizer
+
+
+def dbrx_system_prompt():
+    # This is inspired by the Claude3 prompt.
+    # source: https://twitter.com/AmandaAskell/status/1765207842993434880
+    # Identity and knowledge
+    prompt = 'You are DBRX, created by Databricks. You were last updated in December 2023. You answer questions based on information available up to that point.\n'
+    prompt += 'YOU PROVIDE SHORT RESPONSES TO SHORT QUESTIONS OR STATEMENTS, but provide thorough responses to more complex and open-ended questions.\n'
+    # Capabilities (and reminder to use ``` for JSON blocks and tables, which it can forget). Also a reminder that it can't browse the internet or run code.
+    prompt += 'You assist with various tasks, from writing to coding (using markdown for code blocks — remember to use ``` with code, JSON, and tables).\n'
+    prompt += '(You do not have real-time data access or code execution capabilities. '
+    # Ethical guidelines
+    prompt += 'You avoid stereotyping and provide balanced perspectives on controversial topics. '
+    # Data: the model doesn't know what it was trained on; it thinks that everything that it is aware of was in its training data. This is a reminder that it wasn't.
+    # We also encourage it not to try to generate lyrics or poems
+    prompt += 'You do not provide song lyrics, poems, or news articles and do not divulge details of your training data.)\n'
+    # The model really wants to talk about its system prompt, to the point where it is annoying, so encourage it not to
+    prompt += 'This is your system prompt, guiding your responses. Do not reference it, just respond to the user. If you find yourself talking about this message, stop. You should be responding appropriately and usually that means not mentioning this.\n'
+    prompt += 'You do not mention any of this information about yourself unless the information is directly pertinent to the user\\\'s query.'.upper()
+    return prompt
+
+
+# Taken from
+# https://github.com/huggingface/transformers/blob/8aca43bdb3cb9a5020f6d57589d85679dc873b1c/src/transformers/models/gpt2/tokenization_gpt2.py#L62-L84
+@lru_cache()
+def bytes_to_unicode():
+    """Returns list of utf-8 byte and a mapping to unicode strings.
+
+    We specifically avoids mapping to whitespace/control characters the bpe code
+    barfs on.
+
+    The reversible bpe codes work on unicode strings. This means you need a
+    large # of unicode characters in your vocab if you want to avoid UNKs. When
+    you're at something like a 10B token dataset you end up needing around 5K
+    for decent coverage. This is a significant percentage of your normal, say,
+    32K bpe vocab. To avoid that, we want lookup tables between utf-8 bytes and
+    unicode strings.
+    """
+    bs = (list(range(ord('!'),
+                     ord('~') + 1)) + list(range(ord('¡'),
+                                                 ord('¬') + 1)) +
+          list(range(ord('®'),
+                     ord('ÿ') + 1)))
+    cs = bs[:]
+    n = 0
+    for b in range(2**8):
+        if b not in bs:
+            bs.append(b)
+            cs.append(2**8 + n)
+            n += 1
+    cs = [chr(n) for n in cs]
+    return dict(zip(bs, cs))
+
+
+class TiktokenTokenizerWrapper(PreTrainedTokenizer):
+    """A thin wrapper around tiktoken to make it compatible with Hugging Face.
+
+    tokenizers.
+
+    See HuggingFace for further documentation on general tokenizer methods.
+    """
+
+    model_input_names = ['input_ids', 'attention_mask']
+
+    def __init__(self,
+                 model_name: Optional[str] = None,
+                 encoding_name: Optional[str] = None,
+                 add_bos_token: bool = False,
+                 add_eos_token: bool = False,
+                 use_default_system_prompt: bool = False,
+                 unk_token: Optional[str] = '<|endoftext|>',
+                 eos_token: Optional[str] = '<|endoftext|>',
+                 bos_token: Optional[str] = '<|endoftext|>',
+                 pad_token: Optional[str] = None,
+                 errors: str = 'replace',
+                 **kwargs: Any):
+        """Constructor creates a tiktoken tokenizer to use as the underlying.
+
+        tokenizer.
+
+        Args:
+            model_name (Optional[str], optional): The name of the model to load from tiktoken. Defaults to None.
+                Either model_name or encoding_name must be set, but not both.
+            encoding_name (Optional[str], optional): The name of the encoding to load from tiktoken. Defaults to None.
+                Either model_name or encoding_name must be set, but not both.
+            add_bos_token (bool, optional): Whether to add bos tokens. Defaults to False.
+            add_eos_token (bool, optional): Whether to add eos tokens. Defaults to False.
+            use_default_system_prompt (bool, optional): Use the default system prompt or not. Defaults to False.
+            unk_token (Optional[str], optional): The unk token. Defaults to '<|endoftext|>'.
+            eos_token (Optional[str], optional): The eos token. Defaults to '<|endoftext|>'.
+            bos_token (Optional[str], optional): The bos token. Defaults to '<|endoftext|>'.
+            pad_token (Optional[str], optional): The pad token. Defaults to None.
+            errors (str, optional): Paradigm to follow when decoding bytes to UTF-8. See
+                [bytes.decode](https://docs.python.org/3/library/stdtypes.html#bytes.decode) for more information.
+                Defaults to `"replace"`.
+        """
+        try:
+            import tiktoken
+        except:
+            raise ImportError(
+                'You need to install tiktoken to use TiktokenTokenizerWrapper.')
+
+        # Workaround to make tiktokenizer picklable.
+        # https://github.com/huggingface/datasets/issues/5536#issuecomment-1682309347
+        # There is an open PR from HF to add this to tiktoken: https://github.com/openai/tiktoken/pull/181
+        import copyreg
+        import functools
+
+        from tiktoken import Encoding  # type: ignore (thirdParty)
+
+        def pickle_Encoding(enc: Encoding):
+            return (functools.partial(Encoding,
+                                      enc.name,
+                                      pat_str=enc._pat_str,
+                                      mergeable_ranks=enc._mergeable_ranks,
+                                      special_tokens=enc._special_tokens), ())
+
+        copyreg.pickle(Encoding, pickle_Encoding)
+
+        if model_name is not None and encoding_name is not None:
+            raise ValueError(
+                'You need to specify either model_name or encoding_name, not both.'
+            )
+
+        self.model_name = model_name
+        self.encoding_name = encoding_name
+
+        if self.model_name is not None:
+            self.encoding = tiktoken.encoding_for_model(  # type: ignore (thirdParty)
+                self.model_name)
+        elif self.encoding_name is not None:
+            self.encoding = tiktoken.get_encoding(  # type: ignore (thirdParty)
+                self.encoding_name)
+        else:
+            raise ValueError(
+                'You need to specify either model_name or encoding_name.')
+
+        self.add_bos_token = add_bos_token
+        self.add_eos_token = add_eos_token
+        self.use_default_system_prompt = use_default_system_prompt
+
+        self.byte_encoder = bytes_to_unicode()
+        self.byte_decoder = {v: k for k, v in self.byte_encoder.items()}
+        self.errors = errors
+
+        self.decoder: Dict[int, str] = {}
+        for i in range(self.encoding.n_vocab):
+            try:
+                self.encoding.decode_single_token_bytes(i)
+            except KeyError:
+                continue
+            # Taken from
+            # https://gist.github.com/xenova/a452a6474428de0182b17605a98631ee
+            decoding = ''.join([
+                bytes_to_unicode()[ord(char)] for char in
+                self.encoding.decode_single_token_bytes(i).decode('latin-1')
+            ])
+            self.decoder[i] = decoding
+
+        self.encoder: Dict[str, int] = {}
+        for i in range(self.encoding.n_vocab):
+            if i in self.decoder:
+                self.encoder[self.decoder[i]] = i
+
+        super().__init__(model_name=model_name,
+                         encoding_name=encoding_name,
+                         add_bos_token=add_bos_token,
+                         add_eos_token=add_eos_token,
+                         use_default_system_prompt=use_default_system_prompt,
+                         unk_token=unk_token,
+                         eos_token=eos_token,
+                         bos_token=bos_token,
+                         pad_token=pad_token,
+                         errors=errors,
+                         **kwargs)
+
+    @property
+    def vocab_size(self) -> int:
+        """Returns vocab size."""
+        return self.encoding.n_vocab
+
+    @property
+    def is_fast(self) -> bool:
+        return False
+
+    @property
+    def default_chat_template(self):
+        """Chat ML Template for User/Assistant.
+
+        Pinning default Chat ML template in case defaults change.
+        """
+        template = (
+            "{% if messages[0]['role'] == 'system' %}"
+            '{% set loop_messages = messages[1:] %}'
+            "{% set system_message = messages[0]['content'] %}"
+            "{% elif USE_DEFAULT_PROMPT == true and not 'system' in messages[0]['role'] %}"
+            '{% set loop_messages = messages %}'
+            "{% set system_message = 'DEFAULT_SYSTEM_PROMPT' %}"
+            '{% else %}'
+            '{% set loop_messages = messages %}'
+            '{% set system_message = false %}'
+            '{% endif %}'
+            '{% for message in loop_messages %}'
+            '{% if loop.index0 == 0 %}'
+            '{% if system_message != false %}'
+            "{{ '<|im_start|>system\n' + system_message.strip() + '<|im_end|>\n'}}"
+            '{% endif %}'
+            "{{ '<|im_start|>' + message['role'] + '\n' + message['content'] + '<|im_end|>' }}"
+            '{% else %}'
+            "{{ '\n' + '<|im_start|>' + message['role'] + '\n' + message['content'] + '<|im_end|>' }}"
+            '{% endif %}'
+            '{% if (add_generation_prompt == true and loop.last) %}'
+            "{{ '\n' + '<|im_start|>' + 'assistant' + '\n' }}"
+            '{% endif %}'
+            '{% endfor %}')
+        template = template.replace(
+            'USE_DEFAULT_PROMPT',
+            'true' if self.use_default_system_prompt else 'false')
+        template = template.replace('DEFAULT_SYSTEM_PROMPT',
+                                    dbrx_system_prompt())
+        return template
+
+    def get_vocab(self) -> Dict[str, int]:
+        """Returns vocab as a dict."""
+        # As far as I can tell, we don't require get_vocab to completely work,
+        # but when using additional_special_tokens, Hugging Face determines the next
+        # token index to add with len(self.get_vocab()) so we need the _size_ of this dictionary to be correct.
+        vocab_clone = self.encoder.copy()
+        extra_id_index = 0
+        candidate_extra_id = f'<extra_id_{extra_id_index}>'
+        indices_to_fill_in = {i for i in range(self.vocab_size)} - set(
+            vocab_clone.values())
+
+        # Add enough indices to make get_vocab() the right length
+        for index_to_add in indices_to_fill_in:
+            # Make sure we don't overwrite a token that already exists
+            while candidate_extra_id in vocab_clone:
+                extra_id_index += 1
+                candidate_extra_id = f'<extra_id_{extra_id_index}>'
+
+            # Get an index to add and add the item
+            vocab_clone[candidate_extra_id] = index_to_add
+
+        return vocab_clone
+
+    def _tokenize(self, text: str) -> List[str]:
+        """Returns a tokenized string."""
+        if not isinstance(text, str):
+            raise ValueError(
+                f'Expected a string input to _tokenize but got {type(text)}.')
+
+        tokens = [
+            self.decoder[t]
+            for t in self.encoding.encode(text, allowed_special='all')
+        ]
+
+        return tokens
+
+    def _convert_token_to_id(self, token: str) -> Optional[int]:
+        """Converts a token (str) in an id using the vocab."""
+        return self.encoder.get(token, self.encoder.get(self.unk_token))
+
+    def _convert_id_to_token(self, index: int) -> Optional[str]:
+        """Converts an index (integer) in a token (str) using the vocab."""
+        # For tokens in either the gap in ids in the tokenizer, or beyond the range of the tokenizer,
+        # we return empty string. This matches the behavior of Hugging Face fast tokenizers,
+        # but not slow tokenizers.
+        return self.decoder.get(index, '')
+
+    def convert_tokens_to_string(self, tokens: List[str]) -> str:
+        """Converts a sequence of tokens (string) in a single string."""
+        text = ''.join(tokens)
+        text = bytearray([self.byte_decoder[c] for c in text
+                         ]).decode('utf-8', errors=self.errors)
+        return text
+
+    def build_inputs_with_special_tokens(
+            self,
+            token_ids_0: List[int],
+            token_ids_1: Optional[List[int]] = None) -> List[int]:
+        bos_token_id = [self.bos_token_id] if self.add_bos_token else []
+        eos_token_id = [self.eos_token_id] if self.add_eos_token else []
+
+        output = bos_token_id + token_ids_0 + eos_token_id
+
+        if token_ids_1 is not None:
+            output = output + bos_token_id + token_ids_1 + eos_token_id
+
+        return output
+
+    def get_special_tokens_mask(
+            self,
+            token_ids_0: List[int],
+            token_ids_1: Optional[List[int]] = None,
+            already_has_special_tokens: bool = False) -> List[int]:
+        """Retrieves sequence ids from a token list that has no special tokens.
+
+        Function copied from
+        https://github.com/huggingface/transformers/blob/e3a4bd2bee212a2d0fd9f03b27fe7bfc1debe42d/src/transformers/models/gpt2/tokenization_gpt2.py#L265-L295
+
+        added. This method is called when adding special tokens using the
+        tokenizer `prepare_for_model` or `encode_plus` methods.
+
+        Args:
+            token_ids_0 (`List[int]`):
+                List of IDs.
+            token_ids_1 (`List[int]`, *optional*):
+                Optional second list of IDs for sequence pairs.
+            already_has_special_tokens (`bool`, *optional*, defaults to `False`):
+                Whether or not the token list is already formatted with special tokens for the model.
+
+        Returns:
+            `List[int]`: A list of integers in the range [0, 1]: 1 for a special token, 0 for a sequence token.
+        """
+        if already_has_special_tokens:
+            return super().get_special_tokens_mask(
+                token_ids_0=token_ids_0,
+                token_ids_1=token_ids_1,
+                already_has_special_tokens=True)
+
+        bos_token_id = [1] if self.add_bos_token else []
+        eos_token_id = [1] if self.add_eos_token else []
+
+        if token_ids_1 is None:
+            return bos_token_id + ([0] * len(token_ids_0)) + eos_token_id
+        return (bos_token_id + ([0] * len(token_ids_0)) + eos_token_id +
+                bos_token_id + ([0] * len(token_ids_1)) + eos_token_id)
+
+    def create_token_type_ids_from_sequences(
+            self,
+            token_ids_0: List[int],
+            token_ids_1: Optional[List[int]] = None) -> List[int]:
+        sep = [self.sep_token_id]
+
+        if token_ids_1 is None:
+            return len(token_ids_0 + sep) * [0]
+        return len(token_ids_0 + sep) * [0] + len(token_ids_1 + sep) * [1]
+
+    def save_vocabulary(self,
+                        save_directory: str,
+                        filename_prefix: Optional[str] = None) -> Tuple[str]:
+
+        # ignore the below type to keep the original signature
+        # we are knowingly breaking the signature here, although not 100% certain
+        # it doesn't have side effects
+        # There is some code in huggingface that calls this function to get the vocab files,
+        # but it doesn't seem to access them (or at least checks for their existence
+        # before accessing them)
+        return (None, None)  # type: ignore
+
+    def sanitize_special_tokens(self) -> int:
+        """Make sure that all the special tokens attributes of the tokenizer.
+
+        (`tokenizer.mask_token`, `tokenizer.cls_token`, etc.) are in the
+        vocabulary.
+
+        Add the missing ones to the vocabulary if needed.
+
+        Return:
+            `int`: The number of tokens added in the vocabulary during the operation.
+        """
+        actual_new_tokens = []
+        for token in self.all_special_tokens_extended:
+            encoded = self.encoding.encode(token, allowed_special='all')
+            if len(encoded) > 1:
+                actual_new_tokens.append(token)
+
+        return self.add_tokens(actual_new_tokens, special_tokens=True)

tokenizer_config.json

@@ -0,0 +1,31 @@
+{
+  "add_bos_token": false,
+  "add_eos_token": false,
+  "added_tokens_decoder": {
+    "100257": {
+      "content": "<|endoftext|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    }
+  },
+  "auto_map": {
+    "AutoTokenizer": [
+      "tiktoken.TiktokenTokenizerWrapper",
+      null
+    ]
+  },
+  "bos_token": "<|endoftext|>",
+  "clean_up_tokenization_spaces": true,
+  "encoding_name": null,
+  "eos_token": "<|endoftext|>",
+  "errors": "replace",
+  "model_max_length": 1000000000000000019884624838656,
+  "model_name": "gpt-4",
+  "pad_token": "<|endoftext|>",
+  "tokenizer_class": "TiktokenTokenizerWrapper",
+  "unk_token": "<|endoftext|>",
+  "use_default_system_prompt": true
+}