Add files using upload-large-folder tool

DeepSeek-V3-Base/.gitattributes

@@ -0,0 +1,35 @@
+*.7z filter=lfs diff=lfs merge=lfs -text
+*.arrow filter=lfs diff=lfs merge=lfs -text
+*.bin filter=lfs diff=lfs merge=lfs -text
+*.bz2 filter=lfs diff=lfs merge=lfs -text
+*.ckpt filter=lfs diff=lfs merge=lfs -text
+*.ftz filter=lfs diff=lfs merge=lfs -text
+*.gz filter=lfs diff=lfs merge=lfs -text
+*.h5 filter=lfs diff=lfs merge=lfs -text
+*.joblib filter=lfs diff=lfs merge=lfs -text
+*.lfs.* filter=lfs diff=lfs merge=lfs -text
+*.mlmodel filter=lfs diff=lfs merge=lfs -text
+*.model filter=lfs diff=lfs merge=lfs -text
+*.msgpack filter=lfs diff=lfs merge=lfs -text
+*.npy filter=lfs diff=lfs merge=lfs -text
+*.npz filter=lfs diff=lfs merge=lfs -text
+*.onnx filter=lfs diff=lfs merge=lfs -text
+*.ot filter=lfs diff=lfs merge=lfs -text
+*.parquet filter=lfs diff=lfs merge=lfs -text
+*.pb filter=lfs diff=lfs merge=lfs -text
+*.pickle filter=lfs diff=lfs merge=lfs -text
+*.pkl filter=lfs diff=lfs merge=lfs -text
+*.pt filter=lfs diff=lfs merge=lfs -text
+*.pth filter=lfs diff=lfs merge=lfs -text
+*.rar filter=lfs diff=lfs merge=lfs -text
+*.safetensors filter=lfs diff=lfs merge=lfs -text
+saved_model/**/* filter=lfs diff=lfs merge=lfs -text
+*.tar.* filter=lfs diff=lfs merge=lfs -text
+*.tar filter=lfs diff=lfs merge=lfs -text
+*.tflite filter=lfs diff=lfs merge=lfs -text
+*.tgz filter=lfs diff=lfs merge=lfs -text
+*.wasm filter=lfs diff=lfs merge=lfs -text
+*.xz filter=lfs diff=lfs merge=lfs -text
+*.zip filter=lfs diff=lfs merge=lfs -text
+*.zst filter=lfs diff=lfs merge=lfs -text
+*tfevents* filter=lfs diff=lfs merge=lfs -text

DeepSeek-V3-Base/LICENSE-CODE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2023 DeepSeek
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

DeepSeek-V3-Base/LICENSE-MODEL

@@ -0,0 +1,91 @@
+DEEPSEEK LICENSE AGREEMENT
+
+Version 1.0, 23 October 2023
+
+Copyright (c) 2023 DeepSeek
+
+Section I: PREAMBLE 
+
+Large generative models are being widely adopted and used, and have the potential to transform the way individuals conceive and benefit from AI or ML technologies. 
+
+Notwithstanding the current and potential benefits that these artifacts can bring to society at large, there are also concerns about potential misuses of them, either due to their technical limitations or ethical considerations. 
+
+In short, this license strives for both the open and responsible downstream use of the accompanying model. When it comes to the open character, we took inspiration from open source permissive licenses regarding the grant of IP rights. Referring to the downstream responsible use, we added use-based restrictions not permitting the use of the model in very specific scenarios, in order for the licensor to be able to enforce the license in case potential misuses of the Model may occur. At the same time, we strive to promote open and responsible research on generative models for content generation. 
+
+Even though downstream derivative versions of the model could be released under different licensing terms, the latter will always have to include - at minimum - the same use-based restrictions as the ones in the original license (this license). We believe in the intersection between open and responsible AI development; thus, this agreement aims to strike a balance between both in order to enable responsible open-science in the field of AI. 
+
+This License governs the use of the model (and its derivatives) and is informed by the model card associated with the model. 
+
+NOW THEREFORE, You and DeepSeek agree as follows: 
+
+1. Definitions 
+"License" means the terms and conditions for use, reproduction, and Distribution as defined in this document. 
+"Data" means a collection of information and/or content extracted from the dataset used with the Model, including to train, pretrain, or otherwise evaluate the Model. The Data is not licensed under this License.
+"Output" means the results of operating a Model as embodied in informational content resulting therefrom. 
+"Model" means any accompanying machine-learning based assemblies (including checkpoints), consisting of learnt weights, parameters (including optimizer states), corresponding to the model architecture as embodied in the Complementary Material, that have been trained or tuned, in whole or in part on the Data, using the Complementary Material. 
+"Derivatives of the Model" means all modifications to the Model, works based on the Model, or any other model which is created or initialized by transfer of patterns of the weights, parameters, activations or output of the Model, to the other model, in order to cause the other model to perform similarly to the Model, including - but not limited to - distillation methods entailing the use of intermediate data representations or methods based on the generation of synthetic data by the Model for training the other model. 
+"Complementary Material" means the accompanying source code and scripts used to define, run, load, benchmark or evaluate the Model, and used to prepare data for training or evaluation, if any. This includes any accompanying documentation, tutorials, examples, etc, if any. 
+"Distribution" means any transmission, reproduction, publication or other sharing of the Model or Derivatives of the Model to a third party, including providing the Model as a hosted service made available by electronic or other remote means - e.g. API-based or web access. 
+"DeepSeek" (or "we") means Beijing DeepSeek Artificial Intelligence Fundamental Technology Research Co., Ltd., Hangzhou DeepSeek Artificial Intelligence Fundamental Technology Research Co., Ltd. and/or any of their affiliates.
+"You" (or "Your") means an individual or Legal Entity exercising permissions granted by this License and/or making use of the Model for whichever purpose and in any field of use, including usage of the Model in an end-use application - e.g. chatbot, translator, etc. 
+"Third Parties" means individuals or legal entities that are not under common control with DeepSeek or You. 
+
+Section II: INTELLECTUAL PROPERTY RIGHTS 
+
+Both copyright and patent grants apply to the Model, Derivatives of the Model and Complementary Material. The Model and Derivatives of the Model are subject to additional terms as described in Section III. 
+
+2. Grant of Copyright License. Subject to the terms and conditions of this License, DeepSeek hereby grants to You a perpetual, worldwide, non-exclusive, no-charge, royalty-free, irrevocable copyright license to reproduce, prepare, publicly display, publicly perform, sublicense, and distribute the Complementary Material, the Model, and Derivatives of the Model. 
+
+3. Grant of Patent License. Subject to the terms and conditions of this License and where and as applicable, DeepSeek hereby grants to You a perpetual, worldwide, non-exclusive, no-charge, royalty-free, irrevocable (except as stated in this paragraph) patent license to make, have made, use, offer to sell, sell, import, and otherwise transfer the Model and the Complementary Material, where such license applies only to those patent claims licensable by DeepSeek that are necessarily infringed by its contribution(s). If You institute patent litigation against any entity (including a cross-claim or counterclaim in a lawsuit) alleging that the Model and/or Complementary Material constitutes direct or contributory patent infringement, then any patent licenses granted to You under this License for the Model and/or works shall terminate as of the date such litigation is asserted or filed. 
+
+
+Section III: CONDITIONS OF USAGE, DISTRIBUTION AND REDISTRIBUTION
+
+4. Distribution and Redistribution. You may host for Third Party remote access purposes (e.g. software-as-a-service), reproduce and distribute copies of the Model or Derivatives of the Model thereof in any medium, with or without modifications, provided that You meet the following conditions: 
+a. Use-based restrictions as referenced in paragraph 5 MUST be included as an enforceable provision by You in any type of legal agreement (e.g. a license) governing the use and/or distribution of the Model or Derivatives of the Model, and You shall give notice to subsequent users You Distribute to, that the Model or Derivatives of the Model are subject to paragraph 5. This provision does not apply to the use of Complementary Material. 
+b. You must give any Third Party recipients of the Model or Derivatives of the Model a copy of this License; 
+c. You must cause any modified files to carry prominent notices stating that You changed the files; 
+d. You must retain all copyright, patent, trademark, and attribution notices excluding those notices that do not pertain to any part of the Model, Derivatives of the Model. 
+e. You may add Your own copyright statement to Your modifications and may provide additional or different license terms and conditions - respecting paragraph 4.a. – for use, reproduction, or Distribution of Your modifications, or for any such Derivatives of the Model as a whole, provided Your use, reproduction, and Distribution of the Model otherwise complies with the conditions stated in this License. 
+
+5. Use-based restrictions. The restrictions set forth in Attachment A are considered Use-based restrictions. Therefore You cannot use the Model and the Derivatives of the Model for the specified restricted uses. You may use the Model subject to this License, including only for lawful purposes and in accordance with the License. Use may include creating any content with, finetuning, updating, running, training, evaluating and/or reparametrizing the Model. You shall require all of Your users who use the Model or a Derivative of the Model to comply with the terms of this paragraph (paragraph 5). 
+
+6. The Output You Generate. Except as set forth herein, DeepSeek claims no rights in the Output You generate using the Model. You are accountable for the Output you generate and its subsequent uses. No use of the output can contravene any provision as stated in the License.
+
+Section IV: OTHER PROVISIONS 
+
+7. Updates and Runtime Restrictions. To the maximum extent permitted by law, DeepSeek reserves the right to restrict (remotely or otherwise) usage of the Model in violation of this License. 
+
+8. Trademarks and related. Nothing in this License permits You to make use of DeepSeek’ trademarks, trade names, logos or to otherwise suggest endorsement or misrepresent the relationship between the parties; and any rights not expressly granted herein are reserved by DeepSeek. 
+
+9. Personal information, IP rights and related. This Model may contain personal information and works with IP rights. You commit to complying with applicable laws and regulations in the handling of personal information and the use of such works. Please note that DeepSeek's license granted to you to use the Model does not imply that you have obtained a legitimate basis for processing the related information or works. As an independent personal information processor and IP rights user, you need to ensure full compliance with relevant legal and regulatory requirements when handling personal information and works with IP rights that may be contained in the Model, and are willing to assume solely any risks and consequences that may arise from that.
+
+10. Disclaimer of Warranty. Unless required by applicable law or agreed to in writing, DeepSeek provides the Model and the Complementary Material on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied, including, without limitation, any warranties or conditions of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A PARTICULAR PURPOSE. You are solely responsible for determining the appropriateness of using or redistributing the Model, Derivatives of the Model, and the Complementary Material and assume any risks associated with Your exercise of permissions under this License. 
+
+11. Limitation of Liability. In no event and under no legal theory, whether in tort (including negligence), contract, or otherwise, unless required by applicable law (such as deliberate and grossly negligent acts) or agreed to in writing, shall DeepSeek be liable to You for damages, including any direct, indirect, special, incidental, or consequential damages of any character arising as a result of this License or out of the use or inability to use the Model and the Complementary Material (including but not limited to damages for loss of goodwill, work stoppage, computer failure or malfunction, or any and all other commercial damages or losses), even if DeepSeek has been advised of the possibility of such damages. 
+
+12. Accepting Warranty or Additional Liability. While redistributing the Model, Derivatives of the Model and the Complementary Material thereof, You may choose to offer, and charge a fee for, acceptance of support, warranty, indemnity, or other liability obligations and/or rights consistent with this License. However, in accepting such obligations, You may act only on Your own behalf and on Your sole responsibility, not on behalf of DeepSeek, and only if You agree to indemnify, defend, and hold DeepSeek harmless for any liability incurred by, or claims asserted against, DeepSeek by reason of your accepting any such warranty or additional liability. 
+
+13. If any provision of this License is held to be invalid, illegal or unenforceable, the remaining provisions shall be unaffected thereby and remain valid as if such provision had not been set forth herein.
+
+14. Governing Law and Jurisdiction. This agreement will be governed and construed under PRC laws 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 located in the domicile of Hangzhou DeepSeek Artificial Intelligence Fundamental Technology Research Co., Ltd. shall have exclusive jurisdiction of any dispute arising out of this agreement.
+
+END OF TERMS AND CONDITIONS
+
+Attachment A 
+
+Use Restrictions
+
+You agree not to use the Model or Derivatives of the Model:
+
+-	In any way that violates any applicable national or international law or regulation or infringes upon the lawful rights and interests of any third party; 
+-	For military use in any way;
+-	For the purpose of exploiting, harming or attempting to exploit or harm minors in any way; 
+-	To generate or disseminate verifiably false information and/or content with the purpose of harming others; 
+-	To generate or disseminate inappropriate content subject to applicable regulatory requirements;
+-	To generate or disseminate personal identifiable information without due authorization or for unreasonable use; 
+-	To defame, disparage or otherwise harass others; 
+-	For fully automated decision making that adversely impacts an individual’s legal rights or otherwise creates or modifies a binding, enforceable obligation; 
+-	For any use intended to or which has the effect of discriminating against or harming individuals or groups based on online or offline social behavior or known or predicted personal or personality characteristics; 
+-	To exploit any of the vulnerabilities of a specific group of persons based on their age, social, physical or mental characteristics, in order to materially distort the behavior of a person pertaining to that group in a manner that causes or is likely to cause that person or another person physical or psychological harm; 
+-	For any use intended to or which has the effect of discriminating against individuals or groups based on legally protected characteristics or categories.

DeepSeek-V3-Base/README.md

@@ -0,0 +1,336 @@
+<!-- markdownlint-disable first-line-h1 -->
+<!-- markdownlint-disable html -->
+<!-- markdownlint-disable no-duplicate-header -->
+
+<div align="center">
+  <img src="https://github.com/deepseek-ai/DeepSeek-V2/blob/main/figures/logo.svg?raw=true" width="60%" alt="DeepSeek-V3" />
+</div>
+<hr>
+<div align="center" style="line-height: 1;">
+  <a href="https://www.deepseek.com/" target="_blank" style="margin: 2px;">
+    <img alt="Homepage" src="https://github.com/deepseek-ai/DeepSeek-V2/blob/main/figures/badge.svg?raw=true" style="display: inline-block; vertical-align: middle;"/>
+  </a>
+  <a href="https://chat.deepseek.com/" target="_blank" style="margin: 2px;">
+    <img alt="Chat" src="https://img.shields.io/badge/🤖%20Chat-DeepSeek%20V3-536af5?color=536af5&logoColor=white" style="display: inline-block; vertical-align: middle;"/>
+  </a>
+  <a href="https://huggingface.co/deepseek-ai" target="_blank" style="margin: 2px;">
+    <img alt="Hugging Face" src="https://img.shields.io/badge/%F0%9F%A4%97%20Hugging%20Face-DeepSeek%20AI-ffc107?color=ffc107&logoColor=white" style="display: inline-block; vertical-align: middle;"/>
+  </a>
+</div>
+
+<div align="center" style="line-height: 1;">
+  <a href="https://discord.gg/Tc7c45Zzu5" target="_blank" style="margin: 2px;">
+    <img alt="Discord" src="https://img.shields.io/badge/Discord-DeepSeek%20AI-7289da?logo=discord&logoColor=white&color=7289da" style="display: inline-block; vertical-align: middle;"/>
+  </a>
+  <a href="https://github.com/deepseek-ai/DeepSeek-V2/blob/main/figures/qr.jpeg?raw=true" target="_blank" style="margin: 2px;">
+    <img alt="Wechat" src="https://img.shields.io/badge/WeChat-DeepSeek%20AI-brightgreen?logo=wechat&logoColor=white" style="display: inline-block; vertical-align: middle;"/>
+  </a>
+  <a href="https://twitter.com/deepseek_ai" target="_blank" style="margin: 2px;">
+    <img alt="Twitter Follow" src="https://img.shields.io/badge/Twitter-deepseek_ai-white?logo=x&logoColor=white" style="display: inline-block; vertical-align: middle;"/>
+  </a>
+</div>
+
+<div align="center" style="line-height: 1;">
+  <a href="https://github.com/deepseek-ai/DeepSeek-V3/blob/main/LICENSE-CODE" style="margin: 2px;">
+    <img alt="Code License" src="https://img.shields.io/badge/Code_License-MIT-f5de53?&color=f5de53" style="display: inline-block; vertical-align: middle;"/>
+  </a>
+  <a href="https://github.com/deepseek-ai/DeepSeek-V3/blob/main/LICENSE-MODEL" style="margin: 2px;">
+    <img alt="Model License" src="https://img.shields.io/badge/Model_License-Model_Agreement-f5de53?&color=f5de53" style="display: inline-block; vertical-align: middle;"/>
+  </a>
+</div>
+
+
+<p align="center">
+  <a href="https://github.com/deepseek-ai/DeepSeek-V3/blob/main/DeepSeek_V3.pdf"><b>Paper Link</b>👁️</a>
+</p>
+
+
+## 1. Introduction
+
+We present DeepSeek-V3, a strong Mixture-of-Experts (MoE) language model with 671B total parameters with 37B activated for each token. 
+To achieve efficient inference and cost-effective training, DeepSeek-V3 adopts Multi-head Latent Attention (MLA) and DeepSeekMoE architectures, which were thoroughly validated in DeepSeek-V2. 
+Furthermore, DeepSeek-V3 pioneers an auxiliary-loss-free strategy for load balancing and sets a multi-token prediction training objective for stronger performance. 
+We pre-train DeepSeek-V3 on 14.8 trillion diverse and high-quality tokens, followed by Supervised Fine-Tuning and Reinforcement Learning stages to fully harness its capabilities. 
+Comprehensive evaluations reveal that DeepSeek-V3 outperforms other open-source models and achieves performance comparable to leading closed-source models.
+Despite its excellent performance, DeepSeek-V3 requires only 2.788M H800 GPU hours for its full training.
+In addition, its training process is remarkably stable. 
+Throughout the entire training process, we did not experience any irrecoverable loss spikes or perform any rollbacks. 
+<p align="center">
+  <img width="80%" src="figures/benchmark.png">
+</p>
+
+## 2. Model Summary
+
+---
+
+**Architecture: Innovative Load Balancing Strategy and Training Objective**
+
+- On top of the efficient architecture of DeepSeek-V2, we pioneer an auxiliary-loss-free strategy for load balancing, which minimizes the performance degradation that arises from encouraging load balancing.
+-  We investigate a Multi-Token Prediction (MTP) objective and prove it beneficial to model performance. 
+    It can also be used for speculative decoding for inference acceleration. 
+
+---
+
+**Pre-Training: Towards Ultimate Training Efficiency**
+
+- We design an FP8 mixed precision training framework and, for the first time, validate the feasibility and effectiveness of FP8 training on an extremely large-scale model.  
+- Through co-design of algorithms, frameworks, and hardware, we overcome the communication bottleneck in cross-node MoE training, nearly achieving full computation-communication overlap.  
+  This significantly enhances our training efficiency and reduces the training costs, enabling us to further scale up the model size without additional overhead.  
+- At an economical cost of only 2.664M H800 GPU hours, we complete the pre-training of DeepSeek-V3 on 14.8T tokens, producing the currently strongest open-source base model. The subsequent training stages after pre-training require only 0.1M GPU hours.
+
+---
+
+**Post-Training: Knowledge Distillation from DeepSeek-R1**
+
+-   We introduce an innovative methodology to distill reasoning capabilities from the long-Chain-of-Thought (CoT) model, specifically from one of the DeepSeek R1 series models, into standard LLMs, particularly DeepSeek-V3. Our pipeline elegantly incorporates the verification and reflection patterns of R1 into DeepSeek-V3 and notably improves its reasoning performance. Meanwhile, we also maintain a control over the output style and length of DeepSeek-V3.
+
+---
+
+
+## 3. Model Downloads
+
+<div align="center">
+
+| **Model** | **#Total Params** | **#Activated Params** | **Context Length** | **Download** |
+| :------------: | :------------: | :------------: | :------------: | :------------: |
+| DeepSeek-V3-Base | 671B | 37B | 128K   | [🤗 HuggingFace](https://huggingface.co/deepseek-ai/DeepSeek-V3-Base)   |
+| DeepSeek-V3   | 671B | 37B |  128K   | [🤗 HuggingFace](https://huggingface.co/deepseek-ai/DeepSeek-V3)   |
+
+</div>
+
+**NOTE: The total size of DeepSeek-V3 models on HuggingFace is 685B, which includes 671B of the Main Model weights and 14B of the Multi-Token Prediction (MTP) Module weights.**
+
+To ensure optimal performance and flexibility, we have partnered with open-source communities and hardware vendors to provide multiple ways to run the model locally. For step-by-step guidance, check out Section 6: [How_to Run_Locally](#6-how-to-run-locally).
+
+For developers looking to dive deeper, we recommend exploring [README_WEIGHTS.md](./README_WEIGHTS.md) for details on the Main Model weights and the Multi-Token Prediction (MTP) Modules. Please note that MTP support is currently under active development within the community, and we welcome your contributions and feedback.
+
+## 4. Evaluation Results
+### Base Model
+#### Standard Benchmarks
+
+<div align="center">
+
+
+|  | Benchmark (Metric) | # Shots | DeepSeek-V2 | Qwen2.5 72B | LLaMA3.1 405B | DeepSeek-V3 |
+|---|-------------------|----------|--------|-------------|---------------|---------|
+| | Architecture | - | MoE | Dense | Dense | MoE |
+| | # Activated Params | - | 21B | 72B | 405B | 37B |
+| | # Total Params | - | 236B | 72B | 405B | 671B |
+| English | Pile-test (BPB) | - | 0.606 | 0.638 | **0.542** | 0.548 |
+| | BBH (EM) | 3-shot | 78.8 | 79.8 | 82.9 | **87.5** |
+| | MMLU (Acc.) | 5-shot | 78.4 | 85.0 | 84.4 | **87.1** |
+| | MMLU-Redux (Acc.) | 5-shot | 75.6 | 83.2 | 81.3 | **86.2** |
+| | MMLU-Pro (Acc.) | 5-shot | 51.4 | 58.3 | 52.8 | **64.4** |
+| | DROP (F1) | 3-shot | 80.4 | 80.6 | 86.0 | **89.0** |
+| | ARC-Easy (Acc.) | 25-shot | 97.6 | 98.4 | 98.4 | **98.9** |
+| | ARC-Challenge (Acc.) | 25-shot | 92.2 | 94.5 | **95.3** | **95.3** |
+| | HellaSwag (Acc.) | 10-shot | 87.1 | 84.8 | **89.2** | 88.9 |
+| | PIQA (Acc.) | 0-shot | 83.9 | 82.6 | **85.9** | 84.7 |
+| | WinoGrande (Acc.) | 5-shot | **86.3** | 82.3 | 85.2 | 84.9 |
+| | RACE-Middle (Acc.) | 5-shot | 73.1 | 68.1 | **74.2** | 67.1 |
+| | RACE-High (Acc.) | 5-shot | 52.6 | 50.3 | **56.8** | 51.3 |
+| | TriviaQA (EM) | 5-shot | 80.0 | 71.9 | **82.7** | **82.9** |
+| | NaturalQuestions (EM) | 5-shot | 38.6 | 33.2 | **41.5** | 40.0 |
+| | AGIEval (Acc.) | 0-shot | 57.5 | 75.8 | 60.6 | **79.6** |
+| Code | HumanEval (Pass@1) | 0-shot | 43.3 | 53.0 | 54.9 | **65.2** |
+| | MBPP (Pass@1) | 3-shot | 65.0 | 72.6 | 68.4 | **75.4** |
+| | LiveCodeBench-Base (Pass@1) | 3-shot | 11.6 | 12.9 | 15.5 | **19.4** |
+| | CRUXEval-I (Acc.) | 2-shot | 52.5 | 59.1 | 58.5 | **67.3** |
+| | CRUXEval-O (Acc.) | 2-shot | 49.8 | 59.9 | 59.9 | **69.8** |
+| Math | GSM8K (EM) | 8-shot | 81.6 | 88.3 | 83.5 | **89.3** |
+| | MATH (EM) | 4-shot | 43.4 | 54.4 | 49.0 | **61.6** |
+| | MGSM (EM) | 8-shot | 63.6 | 76.2 | 69.9 | **79.8** |
+| | CMath (EM) | 3-shot | 78.7 | 84.5 | 77.3 | **90.7** |
+| Chinese | CLUEWSC (EM) | 5-shot | 82.0 | 82.5 | **83.0** | 82.7 |
+| | C-Eval (Acc.) | 5-shot | 81.4 | 89.2 | 72.5 | **90.1** |
+| | CMMLU (Acc.) | 5-shot | 84.0 | **89.5** | 73.7 | 88.8 |
+| | CMRC (EM) | 1-shot | **77.4** | 75.8 | 76.0 | 76.3 |
+| | C3 (Acc.) | 0-shot | 77.4 | 76.7 | **79.7** | 78.6 |
+| | CCPM (Acc.) | 0-shot | **93.0** | 88.5 | 78.6 | 92.0 |
+| Multilingual | MMMLU-non-English (Acc.) | 5-shot | 64.0 | 74.8 | 73.8 | **79.4** |
+
+</div>
+
+Note: Best results are shown in bold. Scores with a gap not exceeding 0.3 are considered to be at the same level. DeepSeek-V3 achieves the best performance on most benchmarks, especially on math and code tasks.
+For more evaluation details, please check our paper. 
+
+#### Context Window
+<p align="center">
+  <img width="80%" src="figures/niah.png">
+</p>
+
+Evaluation results on the ``Needle In A Haystack`` (NIAH) tests.  DeepSeek-V3 performs well across all context window lengths up to **128K**. 
+
+### Chat Model
+#### Standard Benchmarks (Models larger than 67B)
+<div align="center">
+
+| | **Benchmark (Metric)** | **DeepSeek V2-0506** | **DeepSeek V2.5-0905** | **Qwen2.5 72B-Inst.** | **Llama3.1 405B-Inst.** | **Claude-3.5-Sonnet-1022** | **GPT-4o 0513** | **DeepSeek V3** |
+|---|---------------------|---------------------|----------------------|---------------------|----------------------|---------------------------|----------------|----------------|
+| | Architecture | MoE | MoE | Dense | Dense | - | - | MoE |
+| | # Activated Params | 21B | 21B | 72B | 405B | - | - | 37B |
+| | # Total Params | 236B | 236B | 72B | 405B | - | - | 671B |
+| English | MMLU (EM) | 78.2 | 80.6 | 85.3 | **88.6** | **88.3** | 87.2 | **88.5** |
+| | MMLU-Redux (EM) | 77.9 | 80.3 | 85.6 | 86.2 | **88.9** | 88.0 | **89.1** |
+| | MMLU-Pro (EM) | 58.5 | 66.2 | 71.6 | 73.3 | **78.0** | 72.6 | 75.9 |
+| | DROP (3-shot F1) | 83.0 | 87.8 | 76.7 | 88.7 | 88.3 | 83.7 | **91.6** |
+| | IF-Eval (Prompt Strict) | 57.7 | 80.6 | 84.1 | 86.0 | **86.5** | 84.3 | 86.1 |
+| | GPQA-Diamond (Pass@1) | 35.3 | 41.3 | 49.0 | 51.1 | **65.0** | 49.9 | 59.1 |
+| | SimpleQA (Correct) | 9.0 | 10.2 | 9.1 | 17.1 | 28.4 | **38.2** | 24.9 |
+| | FRAMES (Acc.) | 66.9 | 65.4 | 69.8 | 70.0 | 72.5 | **80.5** | 73.3 |
+| | LongBench v2 (Acc.) | 31.6 | 35.4 | 39.4 | 36.1 | 41.0 | 48.1 | **48.7** |
+| Code | HumanEval-Mul (Pass@1) | 69.3 | 77.4 | 77.3 | 77.2 | 81.7 | 80.5 | **82.6** |
+| | LiveCodeBench (Pass@1-COT) | 18.8 | 29.2 | 31.1 | 28.4 | 36.3 | 33.4 | **40.5** |
+| | LiveCodeBench (Pass@1) | 20.3 | 28.4 | 28.7 | 30.1 | 32.8 | 34.2 | **37.6** |
+| | Codeforces (Percentile) | 17.5 | 35.6 | 24.8 | 25.3 | 20.3 | 23.6 | **51.6** |
+| | SWE Verified (Resolved) | - | 22.6 | 23.8 | 24.5 | **50.8** | 38.8 | 42.0 |
+| | Aider-Edit (Acc.) | 60.3 | 71.6 | 65.4 | 63.9 | **84.2** | 72.9 | 79.7 |
+| | Aider-Polyglot (Acc.) | - | 18.2 | 7.6 | 5.8 | 45.3 | 16.0 | **49.6** |
+| Math | AIME 2024 (Pass@1) | 4.6 | 16.7 | 23.3 | 23.3 | 16.0 | 9.3 | **39.2** |
+| | MATH-500 (EM) | 56.3 | 74.7 | 80.0 | 73.8 | 78.3 | 74.6 | **90.2** |
+| | CNMO 2024 (Pass@1) | 2.8 | 10.8 | 15.9 | 6.8 | 13.1 | 10.8 | **43.2** |
+| Chinese | CLUEWSC (EM) | 89.9 | 90.4 | **91.4** | 84.7 | 85.4 | 87.9 | 90.9 |
+| | C-Eval (EM) | 78.6 | 79.5 | 86.1 | 61.5 | 76.7 | 76.0 | **86.5** |
+| | C-SimpleQA (Correct) | 48.5 | 54.1 | 48.4 | 50.4 | 51.3 | 59.3 | **64.8** |
+
+Note: All models are evaluated in a configuration that limits the output length to 8K. Benchmarks containing fewer than 1000 samples are tested multiple times using varying temperature settings to derive robust final results. DeepSeek-V3 stands as the best-performing open-source model, and also exhibits competitive performance against frontier closed-source models.
+
+</div>
+
+
+####  Open Ended Generation Evaluation
+
+<div align="center">
+
+
+
+| Model | Arena-Hard | AlpacaEval 2.0 |
+|-------|------------|----------------|
+| DeepSeek-V2.5-0905 | 76.2 | 50.5 |
+| Qwen2.5-72B-Instruct | 81.2 | 49.1 |
+| LLaMA-3.1 405B | 69.3 | 40.5 |
+| GPT-4o-0513 | 80.4 | 51.1 |
+| Claude-Sonnet-3.5-1022 | 85.2 | 52.0 |
+| DeepSeek-V3 | **85.5** | **70.0** |
+
+Note: English open-ended conversation evaluations. For AlpacaEval 2.0, we use the length-controlled win rate as the metric.
+</div>
+
+
+## 5. Chat Website & API Platform
+You can chat with DeepSeek-V3 on DeepSeek's official website: [chat.deepseek.com](https://chat.deepseek.com/sign_in)
+
+We also provide OpenAI-Compatible API at DeepSeek Platform: [platform.deepseek.com](https://platform.deepseek.com/)
+
+## 6. How to Run Locally
+
+DeepSeek-V3 can be deployed locally using the following hardware and open-source community software:
+
+1. **DeepSeek-Infer Demo**: We provide a simple and lightweight demo for FP8 and BF16 inference.
+2. **SGLang**: Fully support the DeepSeek-V3 model in both BF16 and FP8 inference modes.
+3. **LMDeploy**: Enables efficient FP8 and BF16 inference for local and cloud deployment.
+4. **TensorRT-LLM**: Currently supports BF16 inference and INT4/8 quantization, with FP8 support coming soon.
+5. **vLLM**: Support DeekSeek-V3 model with FP8 and BF16 modes for tensor parallelism and pipeline parallelism.
+6. **AMD GPU**: Enables running the DeepSeek-V3 model on AMD GPUs via SGLang in both BF16 and FP8 modes.
+7. **Huawei Ascend NPU**: Supports running DeepSeek-V3 on Huawei Ascend devices.
+
+Since FP8 training is natively adopted in our framework, we only provide FP8 weights. If you require BF16 weights for experimentation, you can use the provided conversion script to perform the transformation.
+
+Here is an example of converting FP8 weights to BF16:
+
+```shell
+cd inference
+python fp8_cast_bf16.py --input-fp8-hf-path /path/to/fp8_weights --output-bf16-hf-path /path/to/bf16_weights
+```
+
+**NOTE: Huggingface's Transformers has not been directly supported yet.**
+
+### 6.1 Inference with DeepSeek-Infer Demo (example only)
+
+#### Model Weights & Demo Code Preparation
+
+First, clone our DeepSeek-V3 GitHub repository:
+
+```shell
+git clone https://github.com/deepseek-ai/DeepSeek-V3.git
+```
+
+Navigate to the `inference` folder and install dependencies listed in `requirements.txt`.
+
+```shell
+cd DeepSeek-V3/inference
+pip install -r requirements.txt
+```
+
+Download the model weights from HuggingFace, and put them into `/path/to/DeepSeek-V3` folder.
+
+#### Model Weights Conversion
+
+Convert HuggingFace model weights to a specific format:
+
+```shell
+python convert.py --hf-ckpt-path /path/to/DeepSeek-V3 --save-path /path/to/DeepSeek-V3-Demo --n-experts 256 --model-parallel 16
+```
+
+#### Run
+
+Then you can chat with DeepSeek-V3:
+
+```shell
+torchrun --nnodes 2 --nproc-per-node 8 generate.py --node-rank $RANK --master-addr $ADDR --ckpt-path /path/to/DeepSeek-V3-Demo --config configs/config_671B.json --interactive --temperature 0.7 --max-new-tokens 200
+```
+
+Or batch inference on a given file:
+
+```shell
+torchrun --nnodes 2 --nproc-per-node 8 generate.py --node-rank $RANK --master-addr $ADDR --ckpt-path /path/to/DeepSeek-V3-Demo --config configs/config_671B.json --input-file $FILE
+```
+
+### 6.2 Inference with SGLang (recommended)
+
+[SGLang](https://github.com/sgl-project/sglang) currently supports MLA optimizations, FP8 (W8A8), FP8 KV Cache, and Torch Compile, delivering state-of-the-art latency and throughput performance among open-source frameworks.
+
+Notably, [SGLang v0.4.1](https://github.com/sgl-project/sglang/releases/tag/v0.4.1) fully supports running DeepSeek-V3 on both **NVIDIA and AMD GPUs**, making it a highly versatile and robust solution.
+
+Here are the launch instructions from the SGLang team: https://github.com/sgl-project/sglang/tree/main/benchmark/deepseek_v3
+
+### 6.3 Inference with LMDeploy (recommended)
+[LMDeploy](https://github.com/InternLM/lmdeploy), a flexible and high-performance inference and serving framework tailored for large language models, now supports DeepSeek-V3. It offers both offline pipeline processing and online deployment capabilities, seamlessly integrating with PyTorch-based workflows.
+
+For comprehensive step-by-step instructions on running DeepSeek-V3 with LMDeploy, please refer to here: https://github.com/InternLM/lmdeploy/issues/2960
+
+
+### 6.4 Inference with TRT-LLM (recommended)
+
+[TensorRT-LLM](https://github.com/NVIDIA/TensorRT-LLM) now supports the DeepSeek-V3 model, offering precision options such as BF16 and INT4/INT8 weight-only. Support for FP8 is currently in progress and will be released soon. You can access the custom branch of TRTLLM specifically for DeepSeek-V3 support through the following link to experience the new features directly: https://github.com/NVIDIA/TensorRT-LLM/tree/deepseek/examples/deepseek_v3. 
+
+### 6.5 Inference with vLLM (recommended)
+
+[vLLM](https://github.com/vllm-project/vllm) v0.6.6 supports DeepSeek-V3 inference for FP8 and BF16 modes on both NVIDIA and AMD GPUs. Aside from standard techniques, vLLM offers _pipeline parallelism_ allowing you to run this model on multiple machines connected by networks. For detailed guidance, please refer to the [vLLM instructions](https://docs.vllm.ai/en/latest/serving/distributed_serving.html). Please feel free to follow [the enhancement plan](https://github.com/vllm-project/vllm/issues/11539) as well.
+
+### 6.6 Recommended Inference Functionality with AMD GPUs
+
+In collaboration with the AMD team, we have achieved Day-One support for AMD GPUs using SGLang, with full compatibility for both FP8 and BF16 precision. For detailed guidance, please refer to the [SGLang instructions](#63-inference-with-lmdeploy-recommended).
+
+### 6.7 Recommended Inference Functionality with Huawei Ascend NPUs
+The [MindIE](https://www.hiascend.com/en/software/mindie) framework from the Huawei Ascend community has successfully adapted the BF16 version of DeepSeek-V3. For step-by-step guidance on Ascend NPUs, please follow the [instructions here](https://modelers.cn/models/MindIE/deepseekv3).
+
+
+## 7. License
+This code repository is licensed under [the MIT License](LICENSE-CODE). The use of DeepSeek-V3 Base/Chat models is subject to [the Model License](LICENSE-MODEL). DeepSeek-V3 series (including Base and Chat) supports commercial use.
+
+## 8. Citation
+```
+@misc{deepseekai2024deepseekv3technicalreport,
+      title={DeepSeek-V3 Technical Report}, 
+      author={DeepSeek-AI},
+      year={2024},
+      eprint={2412.19437},
+      archivePrefix={arXiv},
+      primaryClass={cs.CL},
+      url={https://arxiv.org/abs/2412.19437}, 
+}
+```
+
+## 9. Contact
+If you have any questions, please raise an issue or contact us at [[email protected]]([email protected]).

DeepSeek-V3-Base/README_WEIGHTS.md

@@ -0,0 +1,94 @@
+# DeepSeek-V3 Weight File Documentation
+
+## New Fields in `config.json`
+
+- **model_type**: Specifies the model type, which is updated to `deepseek_v3` in this release.
+- **num_nextn_predict_layers**: Indicates the number of Multi-Token Prediction (MTP) Modules. The open-sourced V3 weights include **1 MTP Module** .
+- **quantization_config**: Describes the configuration for FP8 quantization.
+
+---
+
+## Weight Structure Overview
+
+The DeepSeek-V3 weight file consists of two main components: **Main Model Weights** and **MTP Modules**.
+
+### 1. Main Model Weights
+
+- **Composition**:
+  - Input/output embedding layers and a complete set of 61 Transformer hidden layers.
+- **Parameter Count**:
+  - Total parameters: **671B**
+  - Activation parameters: **36.6B** (including 0.9B for the output Head).
+
+#### Structural Details
+
+- **Embedding Layer**:
+  - `model.embed_tokens.weight`
+- **Transformer Hidden Layers**:
+  - `model.layers.0` to `model.layers.60`, totaling `num_hidden_layers` layers.
+- **Output Layer**:
+  - `model.norm.weight`
+  - `lm_head.weight`
+
+### 2. Multi-Token Prediction (MTP) Modules
+
+- **Composition**:
+  - Additional MTP Modules defined by the `num_nextn_predict_layers` field. In this model, the value is set to 1.
+- **Parameter Count**:
+  - Parameters: **11.5B unique parameters** (excluding the shared 0.9B Embedding and 0.9B output Head).
+  - Activation parameters: **1.5B** (including 0.9B for the output Head).
+
+#### Structural Details
+
+- **embed_tokens**: **Shares parameters** with the Embedding layer of the Main Model weights.
+- **enorm & hnorm**: RMSNorm parameters required for speculative decoding.
+- **eh_proj**: Parameters for dimensionality reduction projection on the norm results.
+- **Additional Transformer Hidden Layer**:
+  - `model.layers.61.self_attn & mlp` (structure identical to the Main Model hidden layers).
+- **shared_head**: **Shares parameters** with the output Head of the Main Model weights.
+
+---
+
+### Loading Rules
+
+- **Main Model Weights**: Loaded via the `num_hidden_layers` parameter in `config.json`.
+- **MTP Modules**: Loaded via the `num_nextn_predict_layers` parameter, with layer IDs appended immediately after the Main Model hidden layers. For example:
+  - If `num_hidden_layers = 61` and `num_nextn_predict_layers = 1`, the MTP Module's layer ID is `61`.
+
+---
+
+## FP8 Weight Documentation
+
+DeepSeek-V3 natively supports FP8 weight format with 128x128 block scaling.
+
+### FP8 Configuration
+
+The FP8 weight file introduces a `quantization_config` field to describe the quantization method. Below is an example configuration:
+
+```json
+"quantization_config": {
+  "activation_scheme": "dynamic",
+  "fmt": "e4m3",
+  "quant_method": "fp8",
+  "weight_block_size": [128, 128]
+}
+```
+
+- **Quantization Format**:
+  - Format type: `fp8` and `e4m3` (corresponding to `torch.float8_e4m3fn`).
+  - Weight block size: `128x128`.
+- **Activation Quantization Scheme**:
+  - Utilizes dynamic activation quantization (`dynamic`).
+
+### Dequantization Method
+
+The FP8 weight file includes a `weight_scale_inv` field, which stores the dequantization scale for each weight block.
+
+- **Storage Format**: `float32 Tensor`, stored alongside the weight data.
+- **Dequantization Formula**:
+  - If the weight block is not aligned to 128, it is zero-padded to 128 before calculating the scale. After quantization, the padded portion is removed.
+  - The dequantization process is performed as: `(128x128 weight block) * weight_scale_inv`.
+
+Through dequantization of the FP8 weights, runtime operations enable online quantization at a granularity of `per-token-per-128-channel`.
+
+---

DeepSeek-V3-Base/configuration_deepseek.py

@@ -0,0 +1,199 @@
+from transformers.configuration_utils import PretrainedConfig
+from transformers.utils import logging
+
+logger = logging.get_logger(__name__)
+
+DEEPSEEK_PRETRAINED_CONFIG_ARCHIVE_MAP = {}
+class DeepseekV3Config(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`DeepseekV3Model`]. It is used to instantiate an DeepSeek
+    model according to the specified arguments, defining the model architecture. Instantiating a configuration with the
+    defaults will yield a similar configuration to that of the DeepSeek-V3.
+
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+
+    Args:
+        vocab_size (`int`, *optional*, defaults to 129280):
+            Vocabulary size of the Deep model. Defines the number of different tokens that can be represented by the
+            `inputs_ids` passed when calling [`DeepseekV3Model`]
+        hidden_size (`int`, *optional*, defaults to 4096):
+            Dimension of the hidden representations.
+        intermediate_size (`int`, *optional*, defaults to 11008):
+            Dimension of the MLP representations.
+        moe_intermediate_size (`int`, *optional*, defaults to 1407):
+            Dimension of the MoE representations.
+        num_hidden_layers (`int`, *optional*, defaults to 32):
+            Number of hidden layers in the Transformer decoder.
+        num_nextn_predict_layers (`int`, *optional*, defaults to 1):
+            Number of nextn predict layers in the DeepSeekV3 Model.
+        num_attention_heads (`int`, *optional*, defaults to 32):
+            Number of attention heads for each attention layer in the Transformer decoder.
+        n_shared_experts (`int`, *optional*, defaults to None):
+            Number of shared experts, None means dense model.
+        n_routed_experts (`int`, *optional*, defaults to None):
+            Number of routed experts, None means dense model.
+        routed_scaling_factor (`float`, *optional*, defaults to 1.0):
+            Scaling factor or routed experts.
+        topk_method (`str`, *optional*, defaults to `gready`):
+            Topk method used in routed gate.
+        n_group (`int`, *optional*, defaults to None):
+            Number of groups for routed experts.
+        topk_group (`int`, *optional*, defaults to None):
+            Number of selected groups for each token(for each token, ensuring the selected experts is only within `topk_group` groups).
+        num_experts_per_tok (`int`, *optional*, defaults to None):
+            Number of selected experts, None means dense model.
+        moe_layer_freq (`int`, *optional*, defaults to 1):
+            The frequency of the MoE layer: one expert layer for every `moe_layer_freq - 1` dense layers.
+        first_k_dense_replace (`int`, *optional*, defaults to 0):
+            Number of dense layers in shallow layers(embed->dense->dense->...->dense->moe->moe...->lm_head).
+                                                            \--k dense layers--/
+        norm_topk_prob (`bool`, *optional*, defaults to False):
+            Whether to normalize the weights of the routed experts.
+        scoring_func (`str`, *optional*, defaults to 'softmax'):
+            Method of computing expert weights.
+        aux_loss_alpha (`float`, *optional*, defaults to 0.001):
+            Auxiliary loss weight coefficient.
+        seq_aux = (`bool`, *optional*, defaults to True):
+            Whether to compute the auxiliary loss for each individual sample.
+        num_key_value_heads (`int`, *optional*):
+            This is the number of key_value heads that should be used to implement Grouped Query Attention. If
+            `num_key_value_heads=num_attention_heads`, the model will use Multi Head Attention (MHA), if
+            `num_key_value_heads=1 the model will use Multi Query Attention (MQA) otherwise GQA is used. When
+            converting a multi-head checkpoint to a GQA checkpoint, each group key and value head should be constructed
+            by meanpooling all the original heads within that group. For more details checkout [this
+            paper](https://arxiv.org/pdf/2305.13245.pdf). If it is not specified, will default to
+            `num_attention_heads`.
+        hidden_act (`str` or `function`, *optional*, defaults to `"silu"`):
+            The non-linear activation function (function or string) in the decoder.
+        max_position_embeddings (`int`, *optional*, defaults to 2048):
+            The maximum sequence length that this model might ever be used with.
+        initializer_range (`float`, *optional*, defaults to 0.02):
+            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+        rms_norm_eps (`float`, *optional*, defaults to 1e-06):
+            The epsilon used by the rms normalization layers.
+        use_cache (`bool`, *optional*, defaults to `True`):
+            Whether or not the model should return the last key/values attentions (not used by all models). Only
+            relevant if `config.is_decoder=True`.
+        pad_token_id (`int`, *optional*):
+            Padding token id.
+        bos_token_id (`int`, *optional*, defaults to 1):
+            Beginning of stream token id.
+        eos_token_id (`int`, *optional*, defaults to 2):
+            End of stream token id.
+        tie_word_embeddings (`bool`, *optional*, defaults to `False`):
+            Whether to tie weight embeddings
+        rope_theta (`float`, *optional*, defaults to 10000.0):
+            The base period of the RoPE embeddings.
+        rope_scaling (`Dict`, *optional*):
+            Dictionary containing the scaling configuration for the RoPE embeddings. Currently supports two scaling
+            strategies: linear and dynamic. Their scaling factor must be a float greater than 1. The expected format is
+            `{"type": strategy name, "factor": scaling factor}`. When using this flag, don't update
+            `max_position_embeddings` to the expected new maximum.
+        attention_bias (`bool`, defaults to `False`, *optional*, defaults to `False`):
+            Whether to use a bias in the query, key, value and output projection layers during self-attention.
+        attention_dropout (`float`, *optional*, defaults to 0.0):
+            The dropout ratio for the attention probabilities.
+
+    ```python
+    >>> from transformers import DeepseekV3Model, DeepseekV3Config
+
+    >>> # Initializing a Deepseek-V3 style configuration
+    >>> configuration = DeepseekV3Config()
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+
+    model_type = "deepseek_v3"
+    keys_to_ignore_at_inference = ["past_key_values"]
+
+    def __init__(
+        self,
+        vocab_size=129280,
+        hidden_size=7168,
+        intermediate_size=18432,
+        moe_intermediate_size = 2048,
+        num_hidden_layers=61,
+        num_nextn_predict_layers=1,
+        num_attention_heads=128,
+        num_key_value_heads=128,
+        n_shared_experts = 1,
+        n_routed_experts = 256,
+        ep_size = 1,
+        routed_scaling_factor = 2.5,
+        kv_lora_rank = 512,
+        q_lora_rank = 1536,
+        qk_rope_head_dim = 64,
+        v_head_dim = 128,
+        qk_nope_head_dim = 128,
+        topk_method = 'noaux_tc',
+        n_group = 8,
+        topk_group = 4,
+        num_experts_per_tok = 8,
+        moe_layer_freq = 1,
+        first_k_dense_replace = 3,
+        norm_topk_prob = True,
+        scoring_func = 'sigmoid',
+        hidden_act="silu",
+        max_position_embeddings=4096,
+        initializer_range=0.02,
+        rms_norm_eps=1e-6,
+        use_cache=True,
+        pad_token_id=None,
+        bos_token_id=0,
+        eos_token_id=1,
+        tie_word_embeddings=False,
+        rope_theta=10000.0,
+        rope_scaling=None,
+        attention_bias=False,
+        attention_dropout=0.0,
+        **kwargs,
+    ):
+        self.vocab_size = vocab_size
+        self.max_position_embeddings = max_position_embeddings
+        self.hidden_size = hidden_size
+        self.intermediate_size = intermediate_size
+        self.moe_intermediate_size = moe_intermediate_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_nextn_predict_layers = num_nextn_predict_layers
+        self.num_attention_heads = num_attention_heads
+        self.n_shared_experts = n_shared_experts
+        self.n_routed_experts = n_routed_experts
+        self.ep_size = ep_size
+        self.routed_scaling_factor = routed_scaling_factor
+        self.kv_lora_rank = kv_lora_rank
+        self.q_lora_rank = q_lora_rank
+        self.qk_rope_head_dim = qk_rope_head_dim
+        self.v_head_dim = v_head_dim
+        self.qk_nope_head_dim = qk_nope_head_dim
+        self.topk_method = topk_method
+        self.n_group = n_group
+        self.topk_group = topk_group
+        self.num_experts_per_tok = num_experts_per_tok
+        self.moe_layer_freq = moe_layer_freq
+        self.first_k_dense_replace = first_k_dense_replace
+        self.norm_topk_prob = norm_topk_prob
+        self.scoring_func = scoring_func
+        # for backward compatibility
+        if num_key_value_heads is None:
+            num_key_value_heads = num_attention_heads
+
+        self.num_key_value_heads = num_key_value_heads
+        self.hidden_act = hidden_act
+        self.initializer_range = initializer_range
+        self.rms_norm_eps = rms_norm_eps
+        self.use_cache = use_cache
+        self.rope_theta = rope_theta
+        self.rope_scaling = rope_scaling
+        self.attention_bias = attention_bias
+        self.attention_dropout = attention_dropout
+
+        super().__init__(
+            pad_token_id=pad_token_id,
+            bos_token_id=bos_token_id,
+            eos_token_id=eos_token_id,
+            tie_word_embeddings=tie_word_embeddings,
+            **kwargs,
+        )

DeepSeek-V3-Base/inference/configs/config_16B.json

@@ -0,0 +1,19 @@
+{
+    "vocab_size": 102400,
+    "dim": 2048,
+    "inter_dim": 10944,
+    "moe_inter_dim": 1408,
+    "n_layers": 27,
+    "n_dense_layers": 1,
+    "n_heads": 16,
+    "n_routed_experts": 64,
+    "n_shared_experts": 2,
+    "n_activated_experts": 6,
+    "route_scale": 1.0,
+    "q_lora_rank": 0,
+    "kv_lora_rank": 512,
+    "qk_nope_head_dim": 128,
+    "qk_rope_head_dim": 64,
+    "v_head_dim": 128,
+    "mscale": 0.707
+}

DeepSeek-V3-Base/inference/configs/config_236B.json

@@ -0,0 +1,20 @@
+{
+    "vocab_size": 102400,
+    "dim": 5120,
+    "inter_dim": 12288,
+    "moe_inter_dim": 1536,
+    "n_layers": 60,
+    "n_dense_layers": 1,
+    "n_heads": 128,
+    "n_routed_experts": 160,
+    "n_shared_experts": 2,
+    "n_activated_experts": 6,
+    "n_expert_groups": 8,
+    "n_limited_groups": 3,
+    "route_scale": 16.0,
+    "q_lora_rank": 1536,
+    "kv_lora_rank": 512,
+    "qk_nope_head_dim": 128,
+    "qk_rope_head_dim": 64,
+    "v_head_dim": 128
+}

DeepSeek-V3-Base/inference/configs/config_671B.json

@@ -0,0 +1,22 @@
+{
+    "vocab_size": 129280,
+    "dim": 7168,
+    "inter_dim": 18432,
+    "moe_inter_dim": 2048,
+    "n_layers": 61,
+    "n_dense_layers": 3,
+    "n_heads": 128,
+    "n_routed_experts": 256,
+    "n_shared_experts": 1,
+    "n_activated_experts": 8,
+    "n_expert_groups": 8,
+    "n_limited_groups": 4,
+    "route_scale": 2.5,
+    "score_func": "sigmoid",
+    "q_lora_rank": 1536,
+    "kv_lora_rank": 512,
+    "qk_nope_head_dim": 128,
+    "qk_rope_head_dim": 64,
+    "v_head_dim": 128,
+    "dtype": "fp8"
+}

DeepSeek-V3-Base/inference/convert.py

@@ -0,0 +1,84 @@
+import os
+import shutil
+from argparse import ArgumentParser
+from glob import glob
+from tqdm import tqdm, trange
+
+import torch
+from safetensors.torch import safe_open, save_file
+
+
+mapping = {
+    "embed_tokens": ("embed", 0),
+    "input_layernorm": ("attn_norm", None),
+    "post_attention_layernorm": ("ffn_norm", None),
+    "q_proj": ("wq", 0),
+    "q_a_proj": ("wq_a", None),
+    "q_a_layernorm": ("q_norm", None),
+    "q_b_proj": ("wq_b", 0),
+    "kv_a_proj_with_mqa": ("wkv_a", None),
+    "kv_a_layernorm": ("kv_norm", None),
+    "kv_b_proj": ("wkv_b", 0),
+    "o_proj": ("wo", 1),
+    "gate": ("gate", None),
+    "gate_proj": ("w1", 0),
+    "down_proj": ("w2", 1),
+    "up_proj": ("w3", 0),
+    "norm": ("norm", None),
+    "lm_head": ("head", 0),
+    "scale": ("scale", None),
+}
+
+
+def main(hf_ckpt_path, save_path, n_experts, mp):
+    torch.set_num_threads(8)
+    n_local_experts = n_experts // mp
+    state_dicts = [{} for _ in range(mp)]
+
+    for file_path in tqdm(glob(os.path.join(hf_ckpt_path, "*.safetensors"))):
+        with safe_open(file_path, framework="pt", device="cpu") as f:
+            for name in f.keys():
+                if "model.layers.61" in name:
+                    continue
+                param: torch.Tensor = f.get_tensor(name)
+                if name.startswith("model."):
+                    name = name[len("model."):]
+                name = name.replace("self_attn", "attn")
+                name = name.replace("mlp", "ffn")
+                name = name.replace("weight_scale_inv", "scale")
+                name = name.replace("e_score_correction_bias", "bias")
+                key = name.split(".")[-2]
+                assert key in mapping
+                new_key, dim = mapping[key]
+                name = name.replace(key, new_key)
+                for i in range(mp):
+                    new_param = param
+                    if "experts" in name and "shared_experts" not in name:
+                        idx = int(name.split(".")[-3])
+                        if idx < i * n_local_experts or idx >= (i + 1) * n_local_experts:
+                            continue
+                    elif dim is not None:
+                        assert param.size(dim) % mp == 0
+                        shard_size = param.size(dim) // mp
+                        new_param = param.narrow(dim, i * shard_size, shard_size).contiguous()
+                    state_dicts[i][name] = new_param
+
+    os.makedirs(save_path, exist_ok=True)
+
+    for i in trange(mp):
+        save_file(state_dicts[i], os.path.join(save_path, f"model{i}-mp{mp}.safetensors"))
+
+    for file_path in glob(os.path.join(hf_ckpt_path, "*token*")):
+        new_file_path = os.path.join(save_path, os.path.basename(file_path))
+        shutil.copyfile(file_path, new_file_path)
+
+
+if __name__ == "__main__":
+    parser = ArgumentParser()
+    parser.add_argument("--hf-ckpt-path", type=str, required=True)
+    parser.add_argument("--save-path", type=str, required=True)
+    parser.add_argument("--n-experts", type=int, required=True)
+    parser.add_argument("--model-parallel", type=int, default=1)
+    args = parser.parse_args()
+    assert args.n_experts % args.model_parallel == 0
+    main(args.hf_ckpt_path, args.save_path, args.n_experts, args.model_parallel)

DeepSeek-V3-Base/inference/fp8_cast_bf16.py

@@ -0,0 +1,81 @@
+import os
+import json
+from argparse import ArgumentParser
+from glob import glob
+from tqdm import tqdm
+
+import torch
+from safetensors.torch import load_file, save_file
+
+from kernel import weight_dequant
+
+def main(fp8_path, bf16_path):
+    torch.set_default_dtype(torch.bfloat16)
+    os.makedirs(bf16_path, exist_ok=True)
+    model_index_file = os.path.join(fp8_path, "model.safetensors.index.json")
+    with open(model_index_file, "r") as f:
+        model_index = json.load(f)
+    weight_map = model_index["weight_map"]
+    
+    # Cache for loaded safetensor files
+    loaded_files = {}
+    fp8_weight_names = []
+
+    # Helper function to get tensor from the correct file
+    def get_tensor(tensor_name):
+        file_name = weight_map[tensor_name]
+        if file_name not in loaded_files:
+            file_path = os.path.join(fp8_path, file_name)
+            loaded_files[file_name] = load_file(file_path, device="cuda")
+        return loaded_files[file_name][tensor_name]
+
+    safetensor_files = list(glob(os.path.join(fp8_path, "*.safetensors")))
+    safetensor_files.sort()
+    for safetensor_file in tqdm(safetensor_files):
+        file_name = os.path.basename(safetensor_file)
+        current_state_dict = load_file(safetensor_file, device="cuda")
+        loaded_files[file_name] = current_state_dict
+        
+        new_state_dict = {}
+        for weight_name, weight in current_state_dict.items():
+            if weight_name.endswith("_scale_inv"):
+                continue
+            elif weight.element_size() == 1:  # FP8 weight
+                scale_inv_name = f"{weight_name}_scale_inv"
+                try:
+                    # Get scale_inv from the correct file
+                    scale_inv = get_tensor(scale_inv_name)
+                    fp8_weight_names.append(weight_name)
+                    new_state_dict[weight_name] = weight_dequant(weight, scale_inv)
+                except KeyError:
+                    print(f"Warning: Missing scale_inv tensor for {weight_name}, skipping conversion")
+                    new_state_dict[weight_name] = weight
+            else:
+                new_state_dict[weight_name] = weight
+                
+        new_safetensor_file = os.path.join(bf16_path, file_name)
+        save_file(new_state_dict, new_safetensor_file)
+        
+        # Memory management: keep only the 2 most recently used files
+        if len(loaded_files) > 2:
+            oldest_file = next(iter(loaded_files))
+            del loaded_files[oldest_file]
+            torch.cuda.empty_cache()
+    
+    # Update model index
+    new_model_index_file = os.path.join(bf16_path, "model.safetensors.index.json")
+    for weight_name in fp8_weight_names:
+        scale_inv_name = f"{weight_name}_scale_inv"
+        if scale_inv_name in weight_map:
+            weight_map.pop(scale_inv_name)
+    with open(new_model_index_file, "w") as f:
+        json.dump({"metadata": {}, "weight_map": weight_map}, f, indent=2)
+        
+
+if __name__ == "__main__":
+    parser = ArgumentParser()
+    parser.add_argument("--input-fp8-hf-path", type=str, required=True)
+    parser.add_argument("--output-bf16-hf-path", type=str, required=True)
+    args = parser.parse_args()
+    main(args.input_fp8_hf_path, args.output_bf16_hf_path)
+

DeepSeek-V3-Base/inference/generate.py

@@ -0,0 +1,137 @@
+import os
+import json
+from argparse import ArgumentParser
+from typing import List
+
+import torch
+import torch.distributed as dist
+from transformers import AutoTokenizer
+from safetensors.torch import load_model
+
+from model import Transformer, ModelArgs
+
+
+def sample(logits, temperature: float = 1.0):
+    logits = logits / max(temperature, 1e-5)
+    probs = torch.softmax(logits, dim=-1)
+    return probs.div_(torch.empty_like(probs).exponential_(1)).argmax(dim=-1)
+
+
+@torch.inference_mode()
+def generate(
+    model: Transformer,
+    prompt_tokens: List[List[int]],
+    max_new_tokens: int,
+    eos_id: int,
+    temperature: float = 1.0
+) -> List[List[int]]:
+    prompt_lens = [len(t) for t in prompt_tokens]
+    assert max(prompt_lens) <= model.max_seq_len
+    total_len = min(model.max_seq_len, max_new_tokens + max(prompt_lens))
+    tokens = torch.full((len(prompt_tokens), total_len), -1, dtype=torch.long, device="cuda")
+    for i, t in enumerate(prompt_tokens):
+        tokens[i, :len(t)] = torch.tensor(t, dtype=torch.long, device="cuda")
+    prev_pos = 0
+    finished = torch.tensor([False] * len(prompt_tokens), device="cuda")
+    prompt_mask = tokens != -1
+    for cur_pos in range(min(prompt_lens), total_len):
+        logits = model.forward(tokens[:, prev_pos:cur_pos], prev_pos)
+        if temperature > 0:
+            next_token = sample(logits, temperature)
+        else:
+            next_token = logits.argmax(dim=-1)
+        next_token = torch.where(prompt_mask[:, cur_pos], tokens[:, cur_pos], next_token)
+        tokens[:, cur_pos] = next_token
+        finished |= torch.logical_and(~prompt_mask[:, cur_pos], next_token == eos_id)
+        prev_pos = cur_pos
+        if finished.all():
+            break
+    completion_tokens = []
+    for i, toks in enumerate(tokens.tolist()):
+        toks = toks[prompt_lens[i]:prompt_lens[i]+max_new_tokens]
+        if eos_id in toks:
+            toks = toks[:toks.index(eos_id)]
+        completion_tokens.append(toks)
+    return completion_tokens
+
+
+def main(
+    ckpt_path: str,
+    config: str,
+    input_file: str = "",
+    interactive: bool = True,
+    max_new_tokens: int = 100,
+    temperature: float = 1.0,
+) -> None:
+    world_size = int(os.getenv("WORLD_SIZE", "1"))
+    rank = int(os.getenv("RANK", "0"))
+    local_rank = int(os.getenv("LOCAL_RANK", "0"))
+    if world_size > 1:
+        dist.init_process_group("nccl")
+    global print
+    if rank != 0:
+        print = lambda *_, **__: None
+    torch.cuda.set_device(local_rank)
+    torch.set_default_dtype(torch.bfloat16)
+    torch.set_num_threads(8)
+    torch.manual_seed(965)
+    with open(config) as f:
+        args = ModelArgs(**json.load(f))
+    print(args)
+    with torch.device("cuda"):
+        model = Transformer(args)
+    tokenizer = AutoTokenizer.from_pretrained(ckpt_path)
+    tokenizer.decode(generate(model, [tokenizer.encode("DeepSeek")], 2, -1, 1.)[0])
+    load_model(model, os.path.join(ckpt_path, f"model{rank}-mp{world_size}.safetensors"))
+
+    if interactive:
+        messages = []
+        while True:
+            if world_size == 1:
+                prompt = input(">>> ")
+            elif rank == 0:
+                prompt = input(">>> ")
+                objects = [prompt]
+                dist.broadcast_object_list(objects, 0)
+            else:
+                objects = [None]
+                dist.broadcast_object_list(objects, 0)
+                prompt = objects[0]
+            if prompt == "/exit":
+                break
+            elif prompt == "/clear":
+                messages.clear()
+                continue
+            messages.append({"role": "user", "content": prompt})
+            prompt_tokens = tokenizer.apply_chat_template(messages, add_generation_prompt=True)
+            completion_tokens = generate(model, [prompt_tokens], max_new_tokens, tokenizer.eos_token_id, temperature)
+            completion = tokenizer.decode(completion_tokens[0], skip_special_tokens=True)
+            print(completion)
+            messages.append({"role": "assistant", "content": completion})
+    else:
+        with open(input_file) as f:
+            prompts = [line.strip() for line in f.readlines()]
+        assert len(prompts) <= args.max_batch_size
+        prompt_tokens = [tokenizer.apply_chat_template([{"role": "user", "content": prompt}], add_generation_prompt=True) for prompt in prompts]
+        completion_tokens = generate(model, prompt_tokens, max_new_tokens, tokenizer.eos_token_id, temperature)
+        completions = tokenizer.batch_decode(completion_tokens, skip_special_tokens=True)
+        for prompt, completion in zip(prompts, completions):
+            print("Prompt:", prompt)
+            print("Completion:", completion)
+            print()
+
+    if world_size > 1:
+        dist.destroy_process_group()
+
+
+if __name__ == "__main__":
+    parser = ArgumentParser()
+    parser.add_argument("--ckpt-path", type=str, required=True)
+    parser.add_argument("--config", type=str, required=True)
+    parser.add_argument("--input-file", type=str, default="")
+    parser.add_argument("--interactive", action="store_true")
+    parser.add_argument("--max-new-tokens", type=int, default=200)
+    parser.add_argument("--temperature", type=float, default=0.2)
+    args = parser.parse_args()
+    assert args.input_file or args.interactive
+    main(args.ckpt_path, args.config, args.input_file, args.interactive, args.max_new_tokens, args.temperature)

DeepSeek-V3-Base/inference/kernel.py

@@ -0,0 +1,108 @@
+from typing import Tuple
+
+import torch
+import triton
+import triton.language as tl
+from triton import Config
+
+
+@triton.jit
+def act_quant_kernel(x_ptr, y_ptr, s_ptr, BLOCK_SIZE: tl.constexpr):
+    pid = tl.program_id(axis=0)
+    offs = pid * BLOCK_SIZE + tl.arange(0, BLOCK_SIZE)
+    x = tl.load(x_ptr + offs).to(tl.float32)
+    s = tl.max(tl.abs(x)) / 448.
+    y = x / s
+    y = y.to(y_ptr.dtype.element_ty)
+    tl.store(y_ptr + offs, y)
+    tl.store(s_ptr + pid, s)
+
+
+def act_quant(x: torch.Tensor, block_size: int = 128) -> Tuple[torch.Tensor, torch.Tensor]:
+    assert x.is_contiguous()
+    assert x.size(-1) % block_size == 0
+    y = torch.empty_like(x, dtype=torch.float8_e4m3fn)
+    s = x.new_empty(*x.size()[:-1], x.size(-1) // block_size, dtype=torch.float32)
+    grid = lambda meta: (triton.cdiv(x.numel(), meta['BLOCK_SIZE']), )
+    act_quant_kernel[grid](x, y, s, BLOCK_SIZE=block_size)
+    return y, s
+
+
+@triton.jit
+def weight_dequant_kernel(x_ptr, s_ptr, y_ptr, M, N, BLOCK_SIZE: tl.constexpr):
+    pid_m = tl.program_id(axis=0)
+    pid_n = tl.program_id(axis=1)
+    n = tl.cdiv(N, BLOCK_SIZE)
+    offs_m = pid_m * BLOCK_SIZE + tl.arange(0, BLOCK_SIZE)
+    offs_n = pid_n * BLOCK_SIZE + tl.arange(0, BLOCK_SIZE)
+    offs = offs_m[:, None] * N + offs_n[None, :]
+    mask = (offs_m[:, None] < M) & (offs_n[None, :] < N)
+    x = tl.load(x_ptr + offs, mask=mask).to(tl.float32)
+    s = tl.load(s_ptr + pid_m * n + pid_n)
+    y = x * s
+    tl.store(y_ptr + offs, y, mask=mask)
+
+
+def weight_dequant(x: torch.Tensor, s: torch.Tensor, block_size: int = 128) -> torch.Tensor:
+    assert x.is_contiguous() and s.is_contiguous()
+    assert x.dim() == 2 and s.dim() == 2
+    M, N = x.size()
+    y = torch.empty_like(x, dtype=torch.get_default_dtype())
+    grid = lambda meta: (triton.cdiv(M, meta['BLOCK_SIZE']), triton.cdiv(N, meta['BLOCK_SIZE']))
+    weight_dequant_kernel[grid](x, s, y, M, N, BLOCK_SIZE=block_size)
+    return y
+
+
+fp8_gemm_configs = [
+    Config({'BLOCK_SIZE_M': block_m, 'BLOCK_SIZE_N': block_n, 'BLOCK_SIZE_K': 128}, num_stages=num_stages, num_warps=8)
+    for block_m in [16, 32, 64] for block_n in [32, 64, 128] for num_stages in [3, 4, 5, 6]
+]
+
+@triton.autotune(configs=fp8_gemm_configs, key=['N', 'K'])
+@triton.jit
+def fp8_gemm_kernel(a_ptr, b_ptr, c_ptr,
+                    a_s_ptr, b_s_ptr,
+                    M, N: tl.constexpr, K: tl.constexpr,
+                    BLOCK_SIZE_M: tl.constexpr,
+                    BLOCK_SIZE_N: tl.constexpr,
+                    BLOCK_SIZE_K: tl.constexpr):
+    pid_m = tl.program_id(axis=0)
+    pid_n = tl.program_id(axis=1)
+    k = tl.cdiv(K, BLOCK_SIZE_K)
+    offs_m = (pid_m * BLOCK_SIZE_M + tl.arange(0, BLOCK_SIZE_M)) % M
+    offs_n = (pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N)) % N
+    offs_k = tl.arange(0, BLOCK_SIZE_K)
+    a_ptrs = a_ptr + offs_m[:, None] * K + offs_k[None, :]
+    b_ptrs = b_ptr + offs_n[None, :] * K + offs_k[:, None]
+    a_s_ptrs = a_s_ptr + offs_m * k
+    b_s_ptrs = b_s_ptr + (offs_n // BLOCK_SIZE_K) * k
+
+    accumulator = tl.zeros((BLOCK_SIZE_M, BLOCK_SIZE_N), dtype=tl.float32)
+    for i in range(k):
+        a = tl.load(a_ptrs, mask=offs_k[None, :] < K - i * BLOCK_SIZE_K, other=0.0)
+        b = tl.load(b_ptrs, mask=offs_k[:, None] < K - i * BLOCK_SIZE_K, other=0.0)
+        a_s = tl.load(a_s_ptrs)
+        b_s = tl.load(b_s_ptrs)
+        accumulator += tl.dot(a, b) * a_s[:, None] * b_s[None, :]
+        a_ptrs += BLOCK_SIZE_K
+        b_ptrs += BLOCK_SIZE_K
+        a_s_ptrs += 1
+        b_s_ptrs += 1
+    c = accumulator.to(c_ptr.dtype.element_ty)
+    offs_m = pid_m * BLOCK_SIZE_M + tl.arange(0, BLOCK_SIZE_M)
+    offs_n = pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N)
+    c_ptrs = c_ptr + offs_m[:, None] * N + offs_n[None, :]
+    mask = (offs_m[:, None] < M) & (offs_n[None, :] < N)
+    tl.store(c_ptrs, c, mask=mask)
+
+
+def fp8_gemm(a: torch.Tensor, a_s: torch.Tensor, b: torch.Tensor, b_s: torch.Tensor):
+    assert a.is_contiguous() and b.is_contiguous()
+    assert a_s.is_contiguous() and b_s.is_contiguous()
+    K = a.size(-1)
+    M = a.numel() // K
+    N = b.size(0)
+    c = a.new_empty(*a.size()[:-1], N, dtype=torch.get_default_dtype())
+    grid = lambda META: (triton.cdiv(M, META['BLOCK_SIZE_M']), triton.cdiv(N, META['BLOCK_SIZE_N']))
+    fp8_gemm_kernel[grid](a, b, c, a_s, b_s, M, N, K)
+    return c

DeepSeek-V3-Base/inference/model.py

@@ -0,0 +1,421 @@
+import math
+from dataclasses import dataclass
+from typing import Tuple, Optional, Literal
+
+import torch
+from torch import nn
+import torch.nn.functional as F
+import torch.distributed as dist
+
+from kernel import act_quant, weight_dequant, fp8_gemm
+
+
+world_size = 1
+rank = 0
+block_size = 128
+gemm_impl: Literal["bf16", "fp8"] = "bf16"
+attn_impl: Literal["naive", "absorb"] = "absorb"
+
+@dataclass
+class ModelArgs:
+    max_batch_size: int = 8
+    max_seq_len: int = 4096 * 4
+    dtype: Literal["bf16", "fp8"] = "bf16"
+    vocab_size: int = 102400
+    dim: int = 2048
+    inter_dim: int = 10944
+    moe_inter_dim: int = 1408
+    n_layers: int = 27
+    n_dense_layers: int = 1
+    n_heads: int = 16
+    # moe
+    n_routed_experts: int = 64
+    n_shared_experts: int = 2
+    n_activated_experts: int = 6
+    n_expert_groups: int = 1
+    n_limited_groups: int = 1
+    score_func: Literal["softmax", "sigmoid"] = "softmax"
+    route_scale: float = 1.
+    # mla
+    q_lora_rank: int = 0
+    kv_lora_rank: int = 512
+    qk_nope_head_dim: int = 128
+    qk_rope_head_dim: int = 64
+    v_head_dim: int = 128
+    # yarn
+    original_seq_len: int = 4096
+    rope_theta: float = 10000.0
+    rope_factor: float = 40
+    beta_fast: int = 32
+    beta_slow: int = 1
+    mscale: float = 1.
+
+
+class ParallelEmbedding(nn.Module):
+    def __init__(self, vocab_size: int, dim: int):
+        super().__init__()
+        self.vocab_size = vocab_size
+        self.dim = dim
+        assert vocab_size % world_size == 0
+        self.part_vocab_size = (vocab_size // world_size)
+        self.vocab_start_idx = rank * self.part_vocab_size
+        self.vocab_end_idx = self.vocab_start_idx + self.part_vocab_size
+        self.weight = nn.Parameter(torch.empty(self.part_vocab_size, self.dim))
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        if world_size > 1:
+            mask = (x < self.vocab_start_idx) | (x >= self.vocab_end_idx)
+            x = x - self.vocab_start_idx
+            x[mask] = 0
+        y = F.embedding(x, self.weight)
+        if world_size > 1:
+            y[mask] = 0
+            dist.all_reduce(y)
+        return y
+
+
+def linear(x: torch.Tensor, weight: torch.Tensor, bias: Optional[torch.Tensor] = None) -> torch.Tensor:
+    if weight.element_size() > 1:
+        return F.linear(x, weight, bias)
+    elif gemm_impl == "bf16":
+        weight = weight_dequant(weight, weight.scale)
+        return F.linear(x, weight, bias)
+    else:
+        x, scale = act_quant(x, block_size)
+        y = fp8_gemm(x, scale, weight, weight.scale)
+        if bias is not None:
+            y += bias
+        return y
+
+
+class Linear(nn.Module):
+    dtype = torch.bfloat16
+
+    def __init__(self, in_features: int, out_features: int, bias: bool = False, dtype = None):
+        super().__init__()
+        self.in_features = in_features
+        self.out_features = out_features
+        self.weight = nn.Parameter(torch.empty(out_features, in_features, dtype=dtype or Linear.dtype))
+        if self.weight.element_size() == 1:
+            scale_out_features = (out_features + block_size - 1) // block_size
+            scale_in_features = (in_features + block_size - 1) // block_size
+            self.weight.scale = self.scale = nn.Parameter(torch.empty(scale_out_features, scale_in_features, dtype=torch.float32))
+        else:
+            self.register_parameter("scale", None)
+        if bias:
+            self.bias = nn.Parameter(torch.empty(self.part_out_features))
+        else:
+            self.register_parameter("bias", None)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        return linear(x, self.weight, self.bias)
+
+
+class ColumnParallelLinear(Linear):
+    def __init__(self, in_features: int, out_features: int, bias: bool = False, dtype = None):
+        assert out_features % world_size == 0
+        self.part_out_features = out_features // world_size
+        super().__init__(in_features, self.part_out_features, bias, dtype)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        y = linear(x, self.weight, self.bias)
+        return y
+
+
+class RowParallelLinear(Linear):
+    def __init__(self, in_features: int, out_features: int, bias: bool = False, dtype = None):
+        assert in_features % world_size == 0
+        self.part_in_features = in_features // world_size
+        super().__init__(self.part_in_features, out_features, bias, dtype)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        y = linear(x, self.weight)
+        if world_size > 1:
+            dist.all_reduce(y)
+        if self.bias is not None:
+            y += self.bias
+        return y
+
+
+class RMSNorm(nn.Module):
+    def __init__(self, dim: int, eps: float = 1e-6):
+        super().__init__()
+        self.eps = eps
+        self.weight = nn.Parameter(torch.ones(dim))
+
+    def forward(self, x: torch.Tensor):
+        x = x.float()
+        y = x * torch.rsqrt(x.pow(2).mean(-1, keepdim=True) + self.eps)
+        return y.type_as(self.weight) * self.weight
+
+
+def precompute_freqs_cis(args: ModelArgs) -> torch.Tensor:
+    dim = args.qk_rope_head_dim
+    seqlen = args.max_seq_len
+    beta_fast = args.beta_fast
+    beta_slow = args.beta_slow
+    base = args.rope_theta
+    factor = args.rope_factor
+
+    def find_correction_dim(num_rotations, dim, base, max_seq_len):
+        return dim * math.log(max_seq_len / (num_rotations * 2 * math.pi)) / (2 * math.log(base))
+
+    def find_correction_range(low_rot, high_rot, dim, base, max_seq_len):
+        low = math.floor(find_correction_dim(low_rot, dim, base, max_seq_len))
+        high = math.ceil(find_correction_dim(high_rot, dim, base, max_seq_len))
+        return max(low, 0), min(high, dim-1)
+
+    def linear_ramp_factor(min, max, dim):
+        if min == max:
+            max += 0.001
+        linear_func = (torch.arange(dim, dtype=torch.float32) - min) / (max - min)
+        ramp_func = torch.clamp(linear_func, 0, 1)
+        return ramp_func
+
+    freqs = 1.0 / (base ** (torch.arange(0, dim, 2, dtype=torch.float32) / dim))
+    if seqlen > args.original_seq_len:
+        low, high = find_correction_range(beta_fast, beta_slow, dim, base, args.original_seq_len)
+        smooth = 1 - linear_ramp_factor(low, high, dim // 2)
+        freqs = freqs / factor * (1 - smooth) + freqs * smooth
+
+    t = torch.arange(seqlen)
+    freqs = torch.outer(t, freqs)
+    freqs_cis = torch.polar(torch.ones_like(freqs), freqs)
+    return freqs_cis
+
+
+def apply_rotary_emb(x: torch.Tensor, freqs_cis: torch.Tensor) -> torch.Tensor:
+    dtype = x.dtype
+    x = torch.view_as_complex(x.float().view(*x.shape[:-1], -1, 2))
+    freqs_cis = freqs_cis.view(1, x.size(1), 1, x.size(-1))
+    y = torch.view_as_real(x * freqs_cis).flatten(3)
+    return y.to(dtype)
+
+
+class MLA(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.dim = args.dim
+        self.n_heads = args.n_heads
+        self.n_local_heads = args.n_heads // world_size
+        self.q_lora_rank = args.q_lora_rank
+        self.kv_lora_rank = args.kv_lora_rank
+        self.qk_nope_head_dim = args.qk_nope_head_dim
+        self.qk_rope_head_dim = args.qk_rope_head_dim
+        self.qk_head_dim = args.qk_nope_head_dim + args.qk_rope_head_dim
+        self.v_head_dim = args.v_head_dim
+
+        if self.q_lora_rank == 0:
+            self.wq = ColumnParallelLinear(self.dim, self.n_heads * self.qk_head_dim)
+        else:
+            self.wq_a = Linear(self.dim, self.q_lora_rank)
+            self.q_norm = RMSNorm(self.q_lora_rank)
+            self.wq_b = ColumnParallelLinear(self.q_lora_rank, self.n_heads * self.qk_head_dim)
+        self.wkv_a = Linear(self.dim, self.kv_lora_rank + self.qk_rope_head_dim)
+        self.kv_norm = RMSNorm(self.kv_lora_rank)
+        self.wkv_b = ColumnParallelLinear(self.kv_lora_rank, self.n_heads * (self.qk_nope_head_dim + self.v_head_dim))
+        self.wo = RowParallelLinear(self.n_heads * self.v_head_dim, self.dim)
+        self.softmax_scale = self.qk_head_dim ** -0.5
+        if args.max_seq_len > args.original_seq_len:
+            mscale = 0.1 * args.mscale * math.log(args.rope_factor) + 1.0
+            self.softmax_scale = self.softmax_scale * mscale * mscale
+
+        if attn_impl == "naive":
+            self.register_buffer("k_cache", torch.zeros(args.max_batch_size, args.max_seq_len, self.n_local_heads, self.qk_head_dim), persistent=False)
+            self.register_buffer("v_cache", torch.zeros(args.max_batch_size, args.max_seq_len, self.n_local_heads, self.v_head_dim), persistent=False)
+        else:
+            self.register_buffer("kv_cache", torch.zeros(args.max_batch_size, args.max_seq_len, self.kv_lora_rank), persistent=False)
+            self.register_buffer("pe_cache", torch.zeros(args.max_batch_size, args.max_seq_len, self.qk_rope_head_dim), persistent=False)
+
+    def forward(self, x: torch.Tensor, start_pos: int, freqs_cis: torch.Tensor, mask: Optional[torch.Tensor]):
+        bsz, seqlen, _ = x.size()
+        end_pos = start_pos + seqlen
+        if self.q_lora_rank == 0:
+            q = self.wq(x)
+        else:
+            q = self.wq_b(self.q_norm(self.wq_a(x)))
+        q = q.view(bsz, seqlen, self.n_local_heads, self.qk_head_dim)
+        q_nope, q_pe = torch.split(q, [self.qk_nope_head_dim, self.qk_rope_head_dim], dim=-1)
+        q_pe = apply_rotary_emb(q_pe, freqs_cis)
+        kv = self.wkv_a(x)
+        kv, k_pe = torch.split(kv, [self.kv_lora_rank, self.qk_rope_head_dim], dim=-1)
+        k_pe = apply_rotary_emb(k_pe.unsqueeze(2), freqs_cis)
+        if attn_impl == "naive":
+            q = torch.cat([q_nope, q_pe], dim=-1)
+            kv = self.wkv_b(self.kv_norm(kv))
+            kv = kv.view(bsz, seqlen, self.n_local_heads, self.qk_nope_head_dim + self.v_head_dim)
+            k_nope, v = torch.split(kv, [self.qk_nope_head_dim, self.v_head_dim], dim=-1)
+            k = torch.cat([k_nope, k_pe.expand(-1, -1, self.n_local_heads, -1)], dim=-1)
+            self.k_cache[:bsz, start_pos:end_pos] = k
+            self.v_cache[:bsz, start_pos:end_pos] = v
+            scores = torch.einsum("bshd,bthd->bsht", q, self.k_cache[:bsz, :end_pos]) * self.softmax_scale
+        else:
+            wkv_b = self.wkv_b.weight if self.wkv_b.scale is None else weight_dequant(self.wkv_b.weight, self.wkv_b.scale, block_size) 
+            wkv_b = wkv_b.view(self.n_local_heads, -1, self.kv_lora_rank)
+            q_nope = torch.einsum("bshd,hdc->bshc", q_nope, wkv_b[:, :self.qk_nope_head_dim])
+            self.kv_cache[:bsz, start_pos:end_pos] = self.kv_norm(kv)
+            self.pe_cache[:bsz, start_pos:end_pos] = k_pe.squeeze(2)
+            scores = (torch.einsum("bshc,btc->bsht", q_nope, self.kv_cache[:bsz, :end_pos]) +
+                      torch.einsum("bshr,btr->bsht", q_pe, self.pe_cache[:bsz, :end_pos])) * self.softmax_scale
+        if mask is not None:
+            scores += mask.unsqueeze(1)
+        scores = scores.softmax(dim=-1, dtype=torch.float32).type_as(x)
+        if attn_impl == "naive":
+            x = torch.einsum("bsht,bthd->bshd", scores, self.v_cache[:bsz, :end_pos])
+        else:
+            x = torch.einsum("bsht,btc->bshc", scores, self.kv_cache[:bsz, :end_pos])
+            x = torch.einsum("bshc,hdc->bshd", x, wkv_b[:, -self.v_head_dim:])
+        x = self.wo(x.flatten(2))
+        return x
+
+
+class MLP(nn.Module):
+    def __init__(self, dim: int, inter_dim: int):
+        super().__init__()
+        self.w1 = ColumnParallelLinear(dim, inter_dim)
+        self.w2 = RowParallelLinear(inter_dim, dim)
+        self.w3 = ColumnParallelLinear(dim, inter_dim)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        return self.w2(F.silu(self.w1(x)) * self.w3(x))
+
+
+class Gate(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.dim = args.dim
+        self.topk = args.n_activated_experts
+        self.n_groups = args.n_expert_groups
+        self.topk_groups = args.n_limited_groups
+        self.score_func = args.score_func
+        self.route_scale = args.route_scale
+        self.weight = nn.Parameter(torch.empty(args.n_routed_experts, args.dim))
+        self.bias = nn.Parameter(torch.empty(args.n_routed_experts)) if self.dim == 7168 else None
+
+    def forward(self, x: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
+        scores = linear(x, self.weight)
+        if self.score_func == "softmax":
+            scores = scores.softmax(dim=-1, dtype=torch.float32)
+        else:
+            scores = scores.sigmoid()
+        original_scores = scores
+        if self.bias is not None:
+            scores = scores + self.bias
+        if self.n_groups > 1:
+            scores = scores.view(x.size(0), self.n_groups, -1)
+            if self.bias is None:
+                group_scores = scores.amax(dim=-1)
+            else:
+                group_scores = scores.topk(2, dim=-1)[0].sum(dim=-1)
+            indices = group_scores.topk(self.topk_groups, dim=-1)[1]
+            mask = torch.zeros_like(scores[..., 0]).scatter_(1, indices, True)
+            scores = (scores * mask.unsqueeze(-1)).flatten(1)
+        indices = torch.topk(scores, self.topk, dim=-1)[1]
+        weights = original_scores.gather(1, indices)
+        if self.score_func == "sigmoid":
+            weights /= weights.sum(dim=-1, keepdim=True)
+        weights *= self.route_scale
+        return weights.type_as(x), indices
+
+
+class Expert(nn.Module):
+    def __init__(self, dim: int, inter_dim: int):
+        super().__init__()
+        self.w1 = Linear(dim, inter_dim)
+        self.w2 = Linear(inter_dim, dim)
+        self.w3 = Linear(dim, inter_dim)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        return self.w2(F.silu(self.w1(x)) * self.w3(x))
+
+
+class MoE(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.dim = args.dim
+        assert args.n_routed_experts % world_size == 0
+        self.n_routed_experts = args.n_routed_experts
+        self.n_local_experts = args.n_routed_experts // world_size
+        self.n_activated_experts = args.n_activated_experts
+        self.experts_start_idx = rank * self.n_local_experts
+        self.experts_end_idx = self.experts_start_idx + self.n_local_experts
+        self.gate = Gate(args)
+        self.experts = nn.ModuleList([Expert(args.dim, args.moe_inter_dim) if self.experts_start_idx <= i < self.experts_end_idx else None
+                                      for i in range(self.n_routed_experts)])
+        self.shared_experts = MLP(args.dim, args.n_shared_experts * args.moe_inter_dim)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        shape = x.size()
+        x = x.view(-1, self.dim)
+        weights, indices = self.gate(x)
+        y = torch.zeros_like(x)
+        counts = torch.bincount(indices.flatten(), minlength=self.n_routed_experts).tolist()
+        for i in range(self.experts_start_idx, self.experts_end_idx):
+            if counts[i] == 0:
+                continue
+            expert = self.experts[i]
+            idx, top = torch.where(indices == i)
+            y[idx] += expert(x[idx]) * weights[idx, top, None]
+        z = self.shared_experts(x)
+        if world_size > 1:
+            dist.all_reduce(y)
+        return (y + z).view(shape)
+
+
+class Block(nn.Module):
+    def __init__(self, layer_id: int, args: ModelArgs):
+        super().__init__()
+        self.attn = MLA(args)
+        self.ffn = MLP(args.dim, args.inter_dim) if layer_id < args.n_dense_layers else MoE(args)
+        self.attn_norm = RMSNorm(args.dim)
+        self.ffn_norm = RMSNorm(args.dim)
+
+    def forward(self, x: torch.Tensor, start_pos: int, freqs_cis: torch.Tensor, mask: Optional[torch.Tensor]) -> torch.Tensor:
+        x = x + self.attn(self.attn_norm(x), start_pos, freqs_cis, mask)
+        x = x + self.ffn(self.ffn_norm(x))
+        return x
+
+
+class Transformer(nn.Module):
+    def __init__(self, args: ModelArgs):
+        global world_size, rank
+        world_size = dist.get_world_size() if dist.is_initialized() else 1
+        rank = dist.get_rank() if dist.is_initialized() else 0
+        Linear.dtype = torch.float8_e4m3fn if args.dtype == "fp8" else torch.bfloat16
+        super().__init__()
+        self.max_seq_len = args.max_seq_len
+        self.embed = ParallelEmbedding(args.vocab_size, args.dim)
+        self.layers = torch.nn.ModuleList()
+        for layer_id in range(args.n_layers):
+            self.layers.append(Block(layer_id, args))
+        self.norm = RMSNorm(args.dim)
+        self.head = ColumnParallelLinear(args.dim, args.vocab_size, dtype=torch.get_default_dtype())
+        self.register_buffer("freqs_cis", precompute_freqs_cis(args), persistent=False)
+
+    @torch.inference_mode()
+    def forward(self, tokens: torch.Tensor, start_pos: int = 0):
+        seqlen = tokens.size(1)
+        h = self.embed(tokens)
+        freqs_cis = self.freqs_cis[start_pos:start_pos+seqlen]
+        mask = None
+        if seqlen > 1:
+            mask = torch.full((seqlen, seqlen), float("-inf"), device=tokens.device).triu_(1)
+        for layer in self.layers:
+            h = layer(h, start_pos, freqs_cis, mask)
+        h = self.norm(h)[:, -1]
+        logits = self.head(h)
+        if world_size > 1:
+            all_logits = [torch.empty_like(logits) for _ in range(world_size)]
+            dist.all_gather(all_logits, logits)
+            logits = torch.cat(all_logits, dim=-1)
+        return logits
+
+
+if __name__ == "__main__":
+    torch.set_default_dtype(torch.bfloat16)
+    torch.set_default_device("cuda")
+    torch.manual_seed(0)
+    args = ModelArgs()
+    x = torch.randint(0, args.vocab_size, (2, 128))
+    model = Transformer(args)
+    print(model(x).size())

DeepSeek-V3-Base/inference/requirements.txt

@@ -0,0 +1,4 @@
+torch==2.4.1
+triton==3.0.0
+transformers==4.46.3
+safetensors==0.4.5

DeepSeek-V3-Base/model-00064-of-000163.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f0885a795d53d6ad311a52e95ceea534001b6d5c286bf289e692bb487f5cb178
+size 4302384963

DeepSeek-V3-Base/model-00069-of-000163.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4253c118e4d644760b84ed9bf6e2548d0a6d4741609ecc6c71cfe9fedef915d8
+size 4302384961

DeepSeek-V3-Base/model-00073-of-000163.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:715341606b17564265e3501b189bcc7e44fa835022150f1a8b01a2607ec5adbe
+size 4302350448

DeepSeek-V3-Base/model-00082-of-000163.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2345f79d6fbac4dfcaddee6c87353b282890c2dfe0d38a964e92c72a63ac607b
+size 4302384518

DeepSeek-V3-Base/model-00084-of-000163.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a6351c0ad501600e343628c13267595c19ae8248358faf3f8f86bceed845635d
+size 4302350602

DeepSeek-V3-Base/model-00085-of-000163.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a666a949436810287696ec5b64e98dcadc313caa02f6c01741f87cbfa397864a
+size 4302384710

DeepSeek-V3-Base/model-00088-of-000163.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:57e28d50941c47e7b4269b27a341e8a151cb60a274983738e05375aff93c3d6e
+size 4302384900

DeepSeek-V3-Base/model-00090-of-000163.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4a7ce55b854ca9b5de72c0d3a9912418d1806deaa04baca3ba2def5e3af69bb6
+size 4302384504

DeepSeek-V3-Base/model-00091-of-000163.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3f12c8d76842664034f68e4d40adec945382d2fb21f745cff110ae3c739854d3
+size 4302384961

DeepSeek-V3-Base/model-00092-of-000163.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6bbfd1f208673e9a262970b5cfd0f1b4393df1f19df80bf9e7af1e8b4f37b0fc
+size 4302350620

DeepSeek-V3-Base/model-00093-of-000163.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:299f42c2a197a4c01a732bfbab61c6e8e49dc32dd5af6fd4650e4119b9855bfd
+size 4302384692

DeepSeek-V3-Base/model-00094-of-000163.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:097a51858b709da50670411463ae20e39ade48218d99c485b7b093002249f059
+size 4302384963

DeepSeek-V3-Base/model-00098-of-000163.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:509289b16b2d717ba671162450803bc961b8cca07b3376f5a471f0e97871d88c
+size 4302384488

DeepSeek-V3-Base/model-00110-of-000163.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6bc5d168729511efd3b80e4fc89b7cd19897ecad3778bd7e5112a8f98d694ddf
+size 4302384900

DeepSeek-V3-Base/model-00116-of-000163.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5a6a24f97970c50f160ef7afae57c340bbb8f293e6f4033282224fac830b40e3
+size 4302384963

DeepSeek-V3-Base/model-00118-of-000163.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8fb725e541e450463fd5a96be319f70ddcbd0214060981005ecaf270d472162f
+size 4302384884

DeepSeek-V3-Base/model-00135-of-000163.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a73e5bc54b533ab7cdfacfc9b5478ab109d45778f87fffc3455a6fbede8814a1
+size 4302384961

DeepSeek-V3-Base/model-00149-of-000163.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2007465688b8a8b2c5fb5f835dbfd98cd6dad012a3242f87183981b46e710320
+size 4302384963

DeepSeek-V3-Base/model-00154-of-000163.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d5544dfa1d120a9666947245537659d51e0e11f000a22bcbdbf20dbb81096b8f
+size 4302384961

flash-attention/.gitignore

@@ -0,0 +1,27 @@
+# Byte-compiled / optimized / DLL files
+__pycache__/
+*.py[cod]
+
+# C extensions
+*.so
+
+# Distribution / packaging
+bin/
+build/
+develop-eggs/
+dist/
+eggs/
+lib/
+lib64/
+parts/
+sdist/
+var/
+*.egg-info/
+.installed.cfg
+*.egg
+
+# IDE-related
+.idea/
+
+# Dev
+venv

flash-attention/.gitmodules

@@ -0,0 +1,3 @@
+[submodule "csrc/cutlass"]
+	path = csrc/cutlass
+	url = https://github.com/NVIDIA/cutlass.git

flash-attention/AUTHORS

@@ -0,0 +1 @@
+Tri Dao, [email protected]

flash-attention/LICENSE

@@ -0,0 +1,29 @@
+BSD 3-Clause License
+
+Copyright (c) 2022, the respective contributors, as shown by the AUTHORS file.
+All rights reserved.
+
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are met:
+
+* Redistributions of source code must retain the above copyright notice, this
+  list of conditions and the following disclaimer.
+
+* Redistributions in binary form must reproduce the above copyright notice,
+  this list of conditions and the following disclaimer in the documentation
+  and/or other materials provided with the distribution.
+
+* Neither the name of the copyright holder nor the names of its
+  contributors may be used to endorse or promote products derived from
+  this software without specific prior written permission.
+
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
+DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
+FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.

flash-attention/MANIFEST.in

@@ -0,0 +1,11 @@
+recursive-include csrc *.cu
+recursive-include csrc *.h
+recursive-include csrc *.cuh
+recursive-include csrc *.cpp
+recursive-include csrc *.hpp
+
+recursive-include flash_attn *.cu
+recursive-include flash_attn *.h
+recursive-include flash_attn *.cuh
+recursive-include flash_attn *.cpp
+recursive-include flash_attn *.hpp

flash-attention/Makefile

@@ -0,0 +1,9 @@
+
+clean_dist:
+	rm -rf dist/*
+
+create_dist: clean_dist
+	python setup.py sdist
+
+upload_package: create_dist
+	twine upload dist/*

flash-attention/README.md

@@ -0,0 +1,417 @@
+# FlashAttention
+This repository provides the official implementation of FlashAttention and
+FlashAttention-2 from the
+following papers.
+
+**FlashAttention: Fast and Memory-Efficient Exact Attention with IO-Awareness**  
+Tri Dao, Daniel Y. Fu, Stefano Ermon, Atri Rudra, Christopher Ré  
+Paper: https://arxiv.org/abs/2205.14135  
+IEEE Spectrum [article](https://spectrum.ieee.org/mlperf-rankings-2022) about our submission to the MLPerf 2.0 benchmark using FlashAttention.
+![FlashAttention](assets/flashattn_banner.jpg)
+
+**FlashAttention-2: Faster Attention with Better Parallelism and Work Partitioning**  
+Tri Dao
+
+Paper: https://tridao.me/publications/flash2/flash2.pdf
+
+![FlashAttention-2](assets/flashattention_logo.png)
+
+
+## Usage
+
+We've been very happy to see FlashAttention being widely adopted in such a short
+time after its release. This [page](https://github.com/Dao-AILab/flash-attention/blob/main/usage.md)
+contains a partial list of places where FlashAttention is being used.
+
+FlashAttention and FlashAttention-2 are free to use and modify (see LICENSE).
+Please cite and credit FlashAttention if you use it.
+
+## Installation and features
+
+Requirements:
+- CUDA 11.6 and above.
+- PyTorch 1.12 and above.
+- Linux. Might work for Windows starting v2.3.2 (we've seen a few positive [reports](https://github.com/Dao-AILab/flash-attention/issues/595)) but Windows compilation still requires more testing. If you have ideas on how to set up prebuilt CUDA wheels for Windows, please reach out via Github issue.
+
+We recommend the
+[Pytorch](https://catalog.ngc.nvidia.com/orgs/nvidia/containers/pytorch)
+container from Nvidia, which has all the required tools to install FlashAttention.
+
+To install:
+1. Make sure that PyTorch is installed.
+2. Make sure that `packaging` is installed (`pip install packaging`)
+3. Make sure that `ninja` is installed and that it works correctly (e.g. `ninja
+--version` then `echo $?` should return exit code 0). If not (sometimes `ninja
+--version` then `echo $?` returns a nonzero exit code), uninstall then reinstall
+`ninja` (`pip uninstall -y ninja && pip install ninja`). Without `ninja`,
+compiling can take a very long time (2h) since it does not use multiple CPU
+cores. With `ninja` compiling takes 3-5 minutes on a 64-core machine.
+4. Then:
+```sh
+pip install flash-attn --no-build-isolation
+```
+Alternatively you can compile from source:
+```sh
+python setup.py install
+```
+
+If your machine has less than 96GB of RAM and lots of CPU cores, `ninja` might
+run too many parallel compilation jobs that could exhaust the amount of RAM. To
+limit the number of parallel compilation jobs, you can set the environment
+variable `MAX_JOBS`:
+```sh
+MAX_JOBS=4 pip install flash-attn --no-build-isolation
+```
+
+Interface: `src/flash_attention_interface.py`
+
+FlashAttention-2 currently supports:
+1. Ampere, Ada, or Hopper GPUs (e.g., A100, RTX 3090, RTX 4090, H100). Support for Turing
+   GPUs (T4, RTX 2080) is coming soon, please use FlashAttention 1.x for Turing
+   GPUs for now.
+2. Datatype fp16 and bf16 (bf16 requires Ampere, Ada, or Hopper GPUs).
+3. All head dimensions up to 256. ~~Head dim > 192 backward requires A100/A800 or H100/H800~~. Head dim 256 backward now works on consumer GPUs (if there's no dropout) as of flash-attn 2.5.5.
+
+
+## How to use FlashAttention
+
+The main functions implement scaled dot product attention (softmax(Q @ K^T *
+softmax_scale) @ V):
+```python
+from flash_attn import flash_attn_qkvpacked_func, flash_attn_func
+```
+
+```python
+flash_attn_qkvpacked_func(qkv, dropout_p=0.0, softmax_scale=None, causal=False,
+                          window_size=(-1, -1), alibi_slopes=None, deterministic=False):
+"""dropout_p should be set to 0.0 during evaluation
+If Q, K, V are already stacked into 1 tensor, this function will be faster than
+calling flash_attn_func on Q, K, V since the backward pass avoids explicit concatenation
+of the gradients of Q, K, V.
+If window_size != (-1, -1), implements sliding window local attention. Query at position i
+will only attend to keys between [i - window_size[0], i + window_size[1]] inclusive.
+Arguments:
+    qkv: (batch_size, seqlen, 3, nheads, headdim)
+    dropout_p: float. Dropout probability.
+    softmax_scale: float. The scaling of QK^T before applying softmax.
+        Default to 1 / sqrt(headdim).
+    causal: bool. Whether to apply causal attention mask (e.g., for auto-regressive modeling).
+    window_size: (left, right). If not (-1, -1), implements sliding window local attention.
+    alibi_slopes: (nheads,) or (batch_size, nheads), fp32. A bias of (-alibi_slope * |i - j|) is added to
+        the attention score of query i and key j.
+    deterministic: bool. Whether to use the deterministic implementation of the backward pass,
+        which is slightly slower and uses more memory. The forward pass is always deterministic.
+Return:
+    out: (batch_size, seqlen, nheads, headdim).
+"""
+```
+
+```python
+flash_attn_func(q, k, v, dropout_p=0.0, softmax_scale=None, causal=False,
+                window_size=(-1, -1), alibi_slopes=None, deterministic=False):
+"""dropout_p should be set to 0.0 during evaluation
+Supports multi-query and grouped-query attention (MQA/GQA) by passing in KV with fewer heads
+than Q. Note that the number of heads in Q must be divisible by the number of heads in KV.
+For example, if Q has 6 heads and K, V have 2 heads, head 0, 1, 2 of Q will attention to head
+0 of K, V, and head 3, 4, 5 of Q will attention to head 1 of K, V.
+If window_size != (-1, -1), implements sliding window local attention. Query at position i
+will only attend to keys between
+[i + seqlen_k - seqlen_q - window_size[0], i + seqlen_k - seqlen_q + window_size[1]] inclusive.
+
+Arguments:
+    q: (batch_size, seqlen, nheads, headdim)
+    k: (batch_size, seqlen, nheads_k, headdim)
+    v: (batch_size, seqlen, nheads_k, headdim)
+    dropout_p: float. Dropout probability.
+    softmax_scale: float. The scaling of QK^T before applying softmax.
+        Default to 1 / sqrt(headdim).
+    causal: bool. Whether to apply causal attention mask (e.g., for auto-regressive modeling).
+    window_size: (left, right). If not (-1, -1), implements sliding window local attention.
+    alibi_slopes: (nheads,) or (batch_size, nheads), fp32. A bias of
+        (-alibi_slope * |i + seqlen_k - seqlen_q - j|)
+        is added to the attention score of query i and key j.
+    deterministic: bool. Whether to use the deterministic implementation of the backward pass,
+        which is slightly slower and uses more memory. The forward pass is always deterministic.
+Return:
+    out: (batch_size, seqlen, nheads, headdim).
+"""
+```
+
+```python
+def flash_attn_with_kvcache(
+    q,
+    k_cache,
+    v_cache,
+    k=None,
+    v=None,
+    rotary_cos=None,
+    rotary_sin=None,
+    cache_seqlens: Optional[Union[(int, torch.Tensor)]] = None,
+    cache_batch_idx: Optional[torch.Tensor] = None,
+    block_table: Optional[torch.Tensor] = None,
+    softmax_scale=None,
+    causal=False,
+    window_size=(-1, -1),  # -1 means infinite context window
+    rotary_interleaved=True,
+    alibi_slopes=None,
+):
+    """
+    If k and v are not None, k_cache and v_cache will be updated *inplace* with the new values from
+    k and v. This is useful for incremental decoding: you can pass in the cached keys/values from
+    the previous step, and update them with the new keys/values from the current step, and do
+    attention with the updated cache, all in 1 kernel.
+
+    If you pass in k / v, you must make sure that the cache is large enough to hold the new values.
+    For example, the KV cache could be pre-allocated with the max sequence length, and you can use
+    cache_seqlens to keep track of the current sequence lengths of each sequence in the batch.
+
+    Also apply rotary embedding if rotary_cos and rotary_sin are passed in. The key @k will be
+    rotated by rotary_cos and rotary_sin at indices cache_seqlens, cache_seqlens + 1, etc.
+    If causal or local (i.e., window_size != (-1, -1)), the query @q will be rotated by rotary_cos
+    and rotary_sin at indices cache_seqlens, cache_seqlens + 1, etc.
+    If not causal and not local, the query @q will be rotated by rotary_cos and rotary_sin at
+    indices cache_seqlens only (i.e. we consider all tokens in @q to be at position cache_seqlens).
+
+    See tests/test_flash_attn.py::test_flash_attn_kvcache for examples of how to use this function.
+
+    Supports multi-query and grouped-query attention (MQA/GQA) by passing in KV with fewer heads
+    than Q. Note that the number of heads in Q must be divisible by the number of heads in KV.
+    For example, if Q has 6 heads and K, V have 2 heads, head 0, 1, 2 of Q will attention to head
+    0 of K, V, and head 3, 4, 5 of Q will attention to head 1 of K, V.
+
+    If causal=True, the causal mask is aligned to the bottom right corner of the attention matrix.
+    For example, if seqlen_q = 2 and seqlen_k = 5, the causal mask (1 = keep, 0 = masked out) is:
+        1 1 1 1 0
+        1 1 1 1 1
+    If seqlen_q = 5 and seqlen_k = 2, the causal mask is:
+        0 0
+        0 0
+        0 0
+        1 0
+        1 1
+    If the row of the mask is all zero, the output will be zero.
+
+    If window_size != (-1, -1), implements sliding window local attention. Query at position i
+    will only attend to keys between
+    [i + seqlen_k - seqlen_q - window_size[0], i + seqlen_k - seqlen_q + window_size[1]] inclusive.
+
+    Note: Does not support backward pass.
+
+    Arguments:
+        q: (batch_size, seqlen, nheads, headdim)
+        k_cache: (batch_size_cache, seqlen_cache, nheads_k, headdim) if there's no block_table,
+            or (num_blocks, page_block_size, nheads_k, headdim) if there's a block_table (i.e. paged KV cache)
+            page_block_size must be a multiple of 256.
+        v_cache: (batch_size_cache, seqlen_cache, nheads_k, headdim) if there's no block_table,
+            or (num_blocks, page_block_size, nheads_k, headdim) if there's a block_table (i.e. paged KV cache)
+        k [optional]: (batch_size, seqlen_new, nheads_k, headdim). If not None, we concatenate
+            k with k_cache, starting at the indices specified by cache_seqlens.
+        v [optional]: (batch_size, seqlen_new, nheads_k, headdim). Similar to k.
+        rotary_cos [optional]: (seqlen_ro, rotary_dim / 2). If not None, we apply rotary embedding
+            to k and q. Only applicable if k and v are passed in. rotary_dim must be divisible by 16.
+        rotary_sin [optional]: (seqlen_ro, rotary_dim / 2). Similar to rotary_cos.
+        cache_seqlens: int, or (batch_size,), dtype torch.int32. The sequence lengths of the
+            KV cache.
+        block_table [optional]: (batch_size, max_num_blocks_per_seq), dtype torch.int32.
+        cache_batch_idx: (batch_size,), dtype torch.int32. The indices used to index into the KV cache.
+            If None, we assume that the batch indices are [0, 1, 2, ..., batch_size - 1].
+            If the indices are not distinct, and k and v are provided, the values updated in the cache
+                 might come from any of the duplicate indices.
+        softmax_scale: float. The scaling of QK^T before applying softmax.
+            Default to 1 / sqrt(headdim).
+        causal: bool. Whether to apply causal attention mask (e.g., for auto-regressive modeling).
+        window_size: (left, right). If not (-1, -1), implements sliding window local attention.
+        rotary_interleaved: bool. Only applicable if rotary_cos and rotary_sin are passed in.
+            If True, rotary embedding will combine dimensions 0 & 1, 2 & 3, etc. If False,
+            rotary embedding will combine dimensions 0 & rotary_dim / 2, 1 & rotary_dim / 2 + 1
+            (i.e. GPT-NeoX style).
+        alibi_slopes: (nheads,) or (batch_size, nheads), fp32. A bias of
+            (-alibi_slope * |i + seqlen_k - seqlen_q - j|)
+            is added to the attention score of query i and key j.
+
+    Return:
+        out: (batch_size, seqlen, nheads, headdim).
+    """
+```
+
+To see how these functions are used in a multi-head attention layer (which
+includes QKV projection, output projection), see the MHA [implementation](https://github.com/Dao-AILab/flash-attention/blob/main/flash_attn/modules/mha.py).
+
+## Changelog
+
+### 2.0: Complete rewrite, 2x faster
+Upgrading from FlashAttention (1.x) to FlashAttention-2
+
+These functions have been renamed:
+- `flash_attn_unpadded_func` -> `flash_attn_varlen_func`
+- `flash_attn_unpadded_qkvpacked_func` -> `flash_attn_varlen_qkvpacked_func`
+- `flash_attn_unpadded_kvpacked_func` -> `flash_attn_varlen_kvpacked_func`
+
+If the inputs have the same sequence lengths in the same batch, it is simpler
+and faster to use these functions:
+```python
+flash_attn_qkvpacked_func(qkv, dropout_p=0.0, softmax_scale=None, causal=False)
+```
+```python
+flash_attn_func(q, k, v, dropout_p=0.0, softmax_scale=None, causal=False)
+```
+### 2.1: Change behavior of causal flag
+
+If seqlen_q != seqlen_k and causal=True, the causal mask is aligned to the
+bottom right corner of the attention matrix, instead of the top-left corner.
+
+For example, if seqlen_q = 2 and seqlen_k = 5, the causal mask (1 = keep, 0 =
+masked out) is:  
+v2.0:  
+    1 0 0 0 0  
+    1 1 0 0 0  
+v2.1:  
+    1 1 1 1 0  
+    1 1 1 1 1  
+
+If seqlen_q = 5 and seqlen_k = 2, the causal mask is:  
+v2.0:  
+    1 0  
+    1 1  
+    1 1  
+    1 1  
+    1 1  
+v2.1:  
+    0 0  
+    0 0  
+    0 0  
+    1 0  
+    1 1  
+If the row of the mask is all zero, the output will be zero.
+
+### 2.2: Optimize for inference
+
+Optimize for inference (iterative decoding) when query has very small sequence
+length (e.g., query sequence length = 1). The bottleneck here is to load KV
+cache as fast as possible, and we split the loading across different thread
+blocks, with a separate kernel to combine results.
+
+See the function `flash_attn_with_kvcache` with more features for inference
+(perform rotary embedding, updating KV cache inplace).
+
+Thanks to the xformers team, and in particular Daniel Haziza, for this
+collaboration.
+
+### 2.3: Local (i.e., sliding window) attention
+
+Implement sliding window attention (i.e., local attention). Thanks to [Mistral
+AI](https://mistral.ai/) and in particular Timothée Lacroix for this
+contribution. Sliding window was used in the [Mistral 7B](https://mistral.ai/news/announcing-mistral-7b/) model.
+
+### 2.4: ALiBi (attention with linear bias), deterministic backward pass.
+
+Implement ALiBi (Press et al., 2021). Thanks to Sanghun Cho from Kakao Brain for this contribution.
+
+Implement deterministic backward pass. Thanks to engineers from [Meituan](www.meituan.com) for this contribution.
+
+### 2.5: Paged KV cache.
+
+Support paged KV cache (i.e., [PagedAttention](https://arxiv.org/abs/2309.06180)).
+Thanks to @beginlner for this contribution.
+
+### 2.6: Softcapping.
+
+Support attention with softcapping, as used in Gemma-2 and Grok models.
+Thanks to @Narsil for this contribution.
+
+## Performance
+
+We present expected speedup (combined forward + backward pass) and memory savings from using FlashAttention against PyTorch standard attention, depending on sequence length, on different GPUs (speedup depends on memory bandwidth - we see more speedup on slower GPU memory).
+
+We currently have benchmarks for these GPUs:
+* [A100](#a100)
+* [H100](#h100)
+<!-- * [RTX 3090](#rtx-3090) -->
+<!-- * [T4](#t4) -->
+
+### A100
+
+We display FlashAttention speedup using these parameters:
+* Head dimension 64 or 128, hidden dimension 2048 (i.e. either 32 or 16 heads).
+* Sequence length 512, 1k, 2k, 4k, 8k, 16k.
+* Batch size set to 16k / seqlen.
+
+#### Speedup
+
+![FlashAttention speedup on A100 80GB SXM5 with FP16/BF16](assets/flash2_a100_fwd_bwd_benchmark.png)
+
+#### Memory
+
+![FlashAttention memory](assets/flashattn_memory.jpg)
+
+We show memory savings in this graph (note that memory footprint is the same no matter if you use dropout or masking).
+Memory savings are proportional to sequence length -- since standard attention has memory quadratic in sequence length, whereas FlashAttention has memory linear in sequence length.
+We see 10X memory savings at sequence length 2K, and 20X at 4K.
+As a result, FlashAttention can scale to much longer sequence lengths.
+
+### H100
+
+![FlashAttention speedup on H100 SXM5 with FP16/BF16](assets/flash2_h100_fwd_bwd_benchmark.png)
+
+## Full model code and training script
+
+We have released the full GPT model
+[implementation](https://github.com/Dao-AILab/flash-attention/blob/main/flash_attn/models/gpt.py).
+We also provide optimized implementations of other layers (e.g., MLP, LayerNorm,
+cross-entropy loss, rotary embedding). Overall this speeds up training by 3-5x
+compared to the baseline implementation from Huggingface, reaching up to 225
+TFLOPs/sec per A100, equivalent to 72% model FLOPs utilization (we don't need
+any activation checkpointing).
+
+We also include a training
+[script](https://github.com/Dao-AILab/flash-attention/tree/main/training) to
+train GPT2 on Openwebtext and GPT3 on The Pile.
+
+## Triton implementation of FlashAttention
+
+Phil Tillet (OpenAI) has an experimental implementation of FlashAttention in Triton:
+https://github.com/openai/triton/blob/master/python/tutorials/06-fused-attention.py
+
+As Triton is a higher-level language than CUDA, it might be easier to understand
+and experiment with. The notations in the Triton implementation are also closer
+to what's used in our paper.
+
+We also have an experimental implementation in Triton that support attention
+bias (e.g. ALiBi):
+https://github.com/Dao-AILab/flash-attention/blob/main/flash_attn/flash_attn_triton.py
+
+
+## Tests
+We test that FlashAttention produces the same output and gradient as a reference
+implementation, up to some numerical tolerance. In particular, we check that the
+maximum numerical error of FlashAttention is at most twice the numerical error
+of a baseline implementation in Pytorch (for different head dimensions, input
+dtype, sequence length, causal / non-causal).
+
+To run the tests:
+```sh
+pytest -q -s tests/test_flash_attn.py
+```
+## When you encounter issues
+
+This new release of FlashAttention-2 has been tested on several GPT-style
+models, mostly on A100 GPUs.
+
+If you encounter bugs, please open a GitHub Issue!
+
+## Citation
+If you use this codebase, or otherwise found our work valuable, please cite:
+```
+@inproceedings{dao2022flashattention,
+  title={Flash{A}ttention: Fast and Memory-Efficient Exact Attention with {IO}-Awareness},
+  author={Dao, Tri and Fu, Daniel Y. and Ermon, Stefano and Rudra, Atri and R{\'e}, Christopher},
+  booktitle={Advances in Neural Information Processing Systems (NeurIPS)},
+  year={2022}
+}
+@inproceedings{dao2023flashattention2,
+  title={Flash{A}ttention-2: Faster Attention with Better Parallelism and Work Partitioning},
+  author={Dao, Tri},
+  booktitle={International Conference on Learning Representations (ICLR)},
+  year={2024}
+}
+```

flash-attention/setup.py

@@ -0,0 +1,386 @@
+# Copyright (c) 2023, Tri Dao.
+
+import sys
+import warnings
+import os
+import re
+import ast
+from pathlib import Path
+from packaging.version import parse, Version
+import platform
+
+from setuptools import setup, find_packages
+import subprocess
+
+import urllib.request
+import urllib.error
+from wheel.bdist_wheel import bdist_wheel as _bdist_wheel
+
+import torch
+from torch.utils.cpp_extension import (
+    BuildExtension,
+    CppExtension,
+    CUDAExtension,
+    CUDA_HOME,
+)
+
+
+with open("README.md", "r", encoding="utf-8") as fh:
+    long_description = fh.read()
+
+
+# ninja build does not work unless include_dirs are abs path
+this_dir = os.path.dirname(os.path.abspath(__file__))
+
+PACKAGE_NAME = "flash_attn"
+
+BASE_WHEEL_URL = (
+    "https://github.com/Dao-AILab/flash-attention/releases/download/{tag_name}/{wheel_name}"
+)
+
+# FORCE_BUILD: Force a fresh build locally, instead of attempting to find prebuilt wheels
+# SKIP_CUDA_BUILD: Intended to allow CI to use a simple `python setup.py sdist` run to copy over raw files, without any cuda compilation
+FORCE_BUILD = os.getenv("FLASH_ATTENTION_FORCE_BUILD", "FALSE") == "TRUE"
+SKIP_CUDA_BUILD = os.getenv("FLASH_ATTENTION_SKIP_CUDA_BUILD", "FALSE") == "TRUE"
+# For CI, we want the option to build with C++11 ABI since the nvcr images use C++11 ABI
+FORCE_CXX11_ABI = os.getenv("FLASH_ATTENTION_FORCE_CXX11_ABI", "FALSE") == "TRUE"
+
+
+def get_platform():
+    """
+    Returns the platform name as used in wheel filenames.
+    """
+    if sys.platform.startswith("linux"):
+        return f'linux_{platform.uname().machine}'
+    elif sys.platform == "darwin":
+        mac_version = ".".join(platform.mac_ver()[0].split(".")[:2])
+        return f"macosx_{mac_version}_x86_64"
+    elif sys.platform == "win32":
+        return "win_amd64"
+    else:
+        raise ValueError("Unsupported platform: {}".format(sys.platform))
+
+
+def get_cuda_bare_metal_version(cuda_dir):
+    raw_output = subprocess.check_output([cuda_dir + "/bin/nvcc", "-V"], universal_newlines=True)
+    output = raw_output.split()
+    release_idx = output.index("release") + 1
+    bare_metal_version = parse(output[release_idx].split(",")[0])
+
+    return raw_output, bare_metal_version
+
+
+def check_if_cuda_home_none(global_option: str) -> None:
+    if CUDA_HOME is not None:
+        return
+    # warn instead of error because user could be downloading prebuilt wheels, so nvcc won't be necessary
+    # in that case.
+    warnings.warn(
+        f"{global_option} was requested, but nvcc was not found.  Are you sure your environment has nvcc available?  "
+        "If you're installing within a container from https://hub.docker.com/r/pytorch/pytorch, "
+        "only images whose names contain 'devel' will provide nvcc."
+    )
+
+
+def append_nvcc_threads(nvcc_extra_args):
+    nvcc_threads = os.getenv("NVCC_THREADS") or "4"
+    return nvcc_extra_args + ["--threads", nvcc_threads]
+
+
+cmdclass = {}
+ext_modules = []
+
+# We want this even if SKIP_CUDA_BUILD because when we run python setup.py sdist we want the .hpp
+# files included in the source distribution, in case the user compiles from source.
+subprocess.run(["git", "submodule", "update", "--init", "csrc/cutlass"])
+
+if not SKIP_CUDA_BUILD:
+    print("\n\ntorch.__version__  = {}\n\n".format(torch.__version__))
+    TORCH_MAJOR = int(torch.__version__.split(".")[0])
+    TORCH_MINOR = int(torch.__version__.split(".")[1])
+
+    # Check, if ATen/CUDAGeneratorImpl.h is found, otherwise use ATen/cuda/CUDAGeneratorImpl.h
+    # See https://github.com/pytorch/pytorch/pull/70650
+    generator_flag = []
+    torch_dir = torch.__path__[0]
+    if os.path.exists(os.path.join(torch_dir, "include", "ATen", "CUDAGeneratorImpl.h")):
+        generator_flag = ["-DOLD_GENERATOR_PATH"]
+
+    check_if_cuda_home_none("flash_attn")
+    # Check, if CUDA11 is installed for compute capability 8.0
+    cc_flag = []
+    if CUDA_HOME is not None:
+        _, bare_metal_version = get_cuda_bare_metal_version(CUDA_HOME)
+        if bare_metal_version < Version("11.6"):
+            raise RuntimeError(
+                "FlashAttention is only supported on CUDA 11.6 and above.  "
+                "Note: make sure nvcc has a supported version by running nvcc -V."
+            )
+    # cc_flag.append("-gencode")
+    # cc_flag.append("arch=compute_75,code=sm_75")
+    cc_flag.append("-gencode")
+    cc_flag.append("arch=compute_80,code=sm_80")
+    if CUDA_HOME is not None:
+        if bare_metal_version >= Version("11.8"):
+            cc_flag.append("-gencode")
+            cc_flag.append("arch=compute_90,code=sm_90")
+
+    # HACK: The compiler flag -D_GLIBCXX_USE_CXX11_ABI is set to be the same as
+    # torch._C._GLIBCXX_USE_CXX11_ABI
+    # https://github.com/pytorch/pytorch/blob/8472c24e3b5b60150096486616d98b7bea01500b/torch/utils/cpp_extension.py#L920
+    if FORCE_CXX11_ABI:
+        torch._C._GLIBCXX_USE_CXX11_ABI = True
+    ext_modules.append(
+        CUDAExtension(
+            name="flash_attn_2_cuda",
+            sources=[
+                "csrc/flash_attn/flash_api.cpp",
+                "csrc/flash_attn/src/flash_fwd_hdim32_fp16_sm80.cu",
+                "csrc/flash_attn/src/flash_fwd_hdim32_bf16_sm80.cu",
+                "csrc/flash_attn/src/flash_fwd_hdim64_fp16_sm80.cu",
+                "csrc/flash_attn/src/flash_fwd_hdim64_bf16_sm80.cu",
+                "csrc/flash_attn/src/flash_fwd_hdim96_fp16_sm80.cu",
+                "csrc/flash_attn/src/flash_fwd_hdim96_bf16_sm80.cu",
+                "csrc/flash_attn/src/flash_fwd_hdim128_fp16_sm80.cu",
+                "csrc/flash_attn/src/flash_fwd_hdim128_bf16_sm80.cu",
+                "csrc/flash_attn/src/flash_fwd_hdim160_fp16_sm80.cu",
+                "csrc/flash_attn/src/flash_fwd_hdim160_bf16_sm80.cu",
+                "csrc/flash_attn/src/flash_fwd_hdim192_fp16_sm80.cu",
+                "csrc/flash_attn/src/flash_fwd_hdim192_bf16_sm80.cu",
+                "csrc/flash_attn/src/flash_fwd_hdim224_fp16_sm80.cu",
+                "csrc/flash_attn/src/flash_fwd_hdim224_bf16_sm80.cu",
+                "csrc/flash_attn/src/flash_fwd_hdim256_fp16_sm80.cu",
+                "csrc/flash_attn/src/flash_fwd_hdim256_bf16_sm80.cu",
+                "csrc/flash_attn/src/flash_fwd_hdim32_fp16_causal_sm80.cu",
+                "csrc/flash_attn/src/flash_fwd_hdim32_bf16_causal_sm80.cu",
+                "csrc/flash_attn/src/flash_fwd_hdim64_fp16_causal_sm80.cu",
+                "csrc/flash_attn/src/flash_fwd_hdim64_bf16_causal_sm80.cu",
+                "csrc/flash_attn/src/flash_fwd_hdim96_fp16_causal_sm80.cu",
+                "csrc/flash_attn/src/flash_fwd_hdim96_bf16_causal_sm80.cu",
+                "csrc/flash_attn/src/flash_fwd_hdim128_fp16_causal_sm80.cu",
+                "csrc/flash_attn/src/flash_fwd_hdim128_bf16_causal_sm80.cu",
+                "csrc/flash_attn/src/flash_fwd_hdim160_fp16_causal_sm80.cu",
+                "csrc/flash_attn/src/flash_fwd_hdim160_bf16_causal_sm80.cu",
+                "csrc/flash_attn/src/flash_fwd_hdim192_fp16_causal_sm80.cu",
+                "csrc/flash_attn/src/flash_fwd_hdim192_bf16_causal_sm80.cu",
+                "csrc/flash_attn/src/flash_fwd_hdim224_fp16_causal_sm80.cu",
+                "csrc/flash_attn/src/flash_fwd_hdim224_bf16_causal_sm80.cu",
+                "csrc/flash_attn/src/flash_fwd_hdim256_fp16_causal_sm80.cu",
+                "csrc/flash_attn/src/flash_fwd_hdim256_bf16_causal_sm80.cu",
+                "csrc/flash_attn/src/flash_bwd_hdim32_fp16_sm80.cu",
+                "csrc/flash_attn/src/flash_bwd_hdim32_bf16_sm80.cu",
+                "csrc/flash_attn/src/flash_bwd_hdim64_fp16_sm80.cu",
+                "csrc/flash_attn/src/flash_bwd_hdim64_bf16_sm80.cu",
+                "csrc/flash_attn/src/flash_bwd_hdim96_fp16_sm80.cu",
+                "csrc/flash_attn/src/flash_bwd_hdim96_bf16_sm80.cu",
+                "csrc/flash_attn/src/flash_bwd_hdim128_fp16_sm80.cu",
+                "csrc/flash_attn/src/flash_bwd_hdim128_bf16_sm80.cu",
+                "csrc/flash_attn/src/flash_bwd_hdim160_fp16_sm80.cu",
+                "csrc/flash_attn/src/flash_bwd_hdim160_bf16_sm80.cu",
+                "csrc/flash_attn/src/flash_bwd_hdim192_fp16_sm80.cu",
+                "csrc/flash_attn/src/flash_bwd_hdim192_bf16_sm80.cu",
+                "csrc/flash_attn/src/flash_bwd_hdim224_fp16_sm80.cu",
+                "csrc/flash_attn/src/flash_bwd_hdim224_bf16_sm80.cu",
+                "csrc/flash_attn/src/flash_bwd_hdim256_fp16_sm80.cu",
+                "csrc/flash_attn/src/flash_bwd_hdim256_bf16_sm80.cu",
+                "csrc/flash_attn/src/flash_fwd_split_hdim32_fp16_sm80.cu",
+                "csrc/flash_attn/src/flash_fwd_split_hdim32_bf16_sm80.cu",
+                "csrc/flash_attn/src/flash_fwd_split_hdim64_fp16_sm80.cu",
+                "csrc/flash_attn/src/flash_fwd_split_hdim64_bf16_sm80.cu",
+                "csrc/flash_attn/src/flash_fwd_split_hdim96_fp16_sm80.cu",
+                "csrc/flash_attn/src/flash_fwd_split_hdim96_bf16_sm80.cu",
+                "csrc/flash_attn/src/flash_fwd_split_hdim128_fp16_sm80.cu",
+                "csrc/flash_attn/src/flash_fwd_split_hdim128_bf16_sm80.cu",
+                "csrc/flash_attn/src/flash_fwd_split_hdim160_fp16_sm80.cu",
+                "csrc/flash_attn/src/flash_fwd_split_hdim160_bf16_sm80.cu",
+                "csrc/flash_attn/src/flash_fwd_split_hdim192_fp16_sm80.cu",
+                "csrc/flash_attn/src/flash_fwd_split_hdim192_bf16_sm80.cu",
+                "csrc/flash_attn/src/flash_fwd_split_hdim224_fp16_sm80.cu",
+                "csrc/flash_attn/src/flash_fwd_split_hdim224_bf16_sm80.cu",
+                "csrc/flash_attn/src/flash_fwd_split_hdim256_fp16_sm80.cu",
+                "csrc/flash_attn/src/flash_fwd_split_hdim256_bf16_sm80.cu",
+                "csrc/flash_attn/src/flash_fwd_split_hdim32_fp16_causal_sm80.cu",
+                "csrc/flash_attn/src/flash_fwd_split_hdim32_bf16_causal_sm80.cu",
+                "csrc/flash_attn/src/flash_fwd_split_hdim64_fp16_causal_sm80.cu",
+                "csrc/flash_attn/src/flash_fwd_split_hdim64_bf16_causal_sm80.cu",
+                "csrc/flash_attn/src/flash_fwd_split_hdim96_fp16_causal_sm80.cu",
+                "csrc/flash_attn/src/flash_fwd_split_hdim96_bf16_causal_sm80.cu",
+                "csrc/flash_attn/src/flash_fwd_split_hdim128_fp16_causal_sm80.cu",
+                "csrc/flash_attn/src/flash_fwd_split_hdim128_bf16_causal_sm80.cu",
+                "csrc/flash_attn/src/flash_fwd_split_hdim160_fp16_causal_sm80.cu",
+                "csrc/flash_attn/src/flash_fwd_split_hdim160_bf16_causal_sm80.cu",
+                "csrc/flash_attn/src/flash_fwd_split_hdim192_fp16_causal_sm80.cu",
+                "csrc/flash_attn/src/flash_fwd_split_hdim192_bf16_causal_sm80.cu",
+                "csrc/flash_attn/src/flash_fwd_split_hdim224_fp16_causal_sm80.cu",
+                "csrc/flash_attn/src/flash_fwd_split_hdim224_bf16_causal_sm80.cu",
+                "csrc/flash_attn/src/flash_fwd_split_hdim256_fp16_causal_sm80.cu",
+                "csrc/flash_attn/src/flash_fwd_split_hdim256_bf16_causal_sm80.cu",
+            ],
+            extra_compile_args={
+                "cxx": ["-O3", "-std=c++17"] + generator_flag,
+                "nvcc": append_nvcc_threads(
+                    [
+                        "-O3",
+                        "-std=c++17",
+                        "-U__CUDA_NO_HALF_OPERATORS__",
+                        "-U__CUDA_NO_HALF_CONVERSIONS__",
+                        "-U__CUDA_NO_HALF2_OPERATORS__",
+                        "-U__CUDA_NO_BFLOAT16_CONVERSIONS__",
+                        "--expt-relaxed-constexpr",
+                        "--expt-extended-lambda",
+                        "--use_fast_math",
+                        # "--ptxas-options=-v",
+                        # "--ptxas-options=-O2",
+                        # "-lineinfo",
+                        # "-DFLASHATTENTION_DISABLE_BACKWARD",
+                        # "-DFLASHATTENTION_DISABLE_DROPOUT",
+                        # "-DFLASHATTENTION_DISABLE_ALIBI",
+                        # "-DFLASHATTENTION_DISABLE_SOFTCAP",
+                        # "-DFLASHATTENTION_DISABLE_UNEVEN_K",
+                        # "-DFLASHATTENTION_DISABLE_LOCAL",
+                    ]
+                    + generator_flag
+                    + cc_flag
+                ),
+            },
+            include_dirs=[
+                Path(this_dir) / "csrc" / "flash_attn",
+                Path(this_dir) / "csrc" / "flash_attn" / "src",
+                Path(this_dir) / "csrc" / "cutlass" / "include",
+            ],
+        )
+    )
+
+
+def get_package_version():
+    with open(Path(this_dir) / "flash_attn" / "__init__.py", "r") as f:
+        version_match = re.search(r"^__version__\s*=\s*(.*)$", f.read(), re.MULTILINE)
+    public_version = ast.literal_eval(version_match.group(1))
+    local_version = os.environ.get("FLASH_ATTN_LOCAL_VERSION")
+    if local_version:
+        return f"{public_version}+{local_version}"
+    else:
+        return str(public_version)
+
+
+def get_wheel_url():
+    # Determine the version numbers that will be used to determine the correct wheel
+    # We're using the CUDA version used to build torch, not the one currently installed
+    # _, cuda_version_raw = get_cuda_bare_metal_version(CUDA_HOME)
+    torch_cuda_version = parse(torch.version.cuda)
+    torch_version_raw = parse(torch.__version__)
+    # For CUDA 11, we only compile for CUDA 11.8, and for CUDA 12 we only compile for CUDA 12.2
+    # to save CI time. Minor versions should be compatible.
+    torch_cuda_version = parse("11.8") if torch_cuda_version.major == 11 else parse("12.2")
+    python_version = f"cp{sys.version_info.major}{sys.version_info.minor}"
+    platform_name = get_platform()
+    flash_version = get_package_version()
+    # cuda_version = f"{cuda_version_raw.major}{cuda_version_raw.minor}"
+    cuda_version = f"{torch_cuda_version.major}{torch_cuda_version.minor}"
+    torch_version = f"{torch_version_raw.major}.{torch_version_raw.minor}"
+    cxx11_abi = str(torch._C._GLIBCXX_USE_CXX11_ABI).upper()
+
+    # Determine wheel URL based on CUDA version, torch version, python version and OS
+    wheel_filename = f"{PACKAGE_NAME}-{flash_version}+cu{cuda_version}torch{torch_version}cxx11abi{cxx11_abi}-{python_version}-{python_version}-{platform_name}.whl"
+    wheel_url = BASE_WHEEL_URL.format(tag_name=f"v{flash_version}", wheel_name=wheel_filename)
+    return wheel_url, wheel_filename
+
+
+class CachedWheelsCommand(_bdist_wheel):
+    """
+    The CachedWheelsCommand plugs into the default bdist wheel, which is ran by pip when it cannot
+    find an existing wheel (which is currently the case for all flash attention installs). We use
+    the environment parameters to detect whether there is already a pre-built version of a compatible
+    wheel available and short-circuits the standard full build pipeline.
+    """
+
+    def run(self):
+        if FORCE_BUILD:
+            return super().run()
+
+        wheel_url, wheel_filename = get_wheel_url()
+        print("Guessing wheel URL: ", wheel_url)
+        try:
+            urllib.request.urlretrieve(wheel_url, wheel_filename)
+
+            # Make the archive
+            # Lifted from the root wheel processing command
+            # https://github.com/pypa/wheel/blob/cf71108ff9f6ffc36978069acb28824b44ae028e/src/wheel/bdist_wheel.py#LL381C9-L381C85
+            if not os.path.exists(self.dist_dir):
+                os.makedirs(self.dist_dir)
+
+            impl_tag, abi_tag, plat_tag = self.get_tag()
+            archive_basename = f"{self.wheel_dist_name}-{impl_tag}-{abi_tag}-{plat_tag}"
+
+            wheel_path = os.path.join(self.dist_dir, archive_basename + ".whl")
+            print("Raw wheel path", wheel_path)
+            os.rename(wheel_filename, wheel_path)
+        except (urllib.error.HTTPError, urllib.error.URLError):
+            print("Precompiled wheel not found. Building from source...")
+            # If the wheel could not be downloaded, build from source
+            super().run()
+
+
+class NinjaBuildExtension(BuildExtension):
+    def __init__(self, *args, **kwargs) -> None:
+        # do not override env MAX_JOBS if already exists
+        if not os.environ.get("MAX_JOBS"):
+            import psutil
+
+            # calculate the maximum allowed NUM_JOBS based on cores
+            max_num_jobs_cores = max(1, os.cpu_count() // 2)
+
+            # calculate the maximum allowed NUM_JOBS based on free memory
+            free_memory_gb = psutil.virtual_memory().available / (1024 ** 3)  # free memory in GB
+            max_num_jobs_memory = int(free_memory_gb / 9)  # each JOB peak memory cost is ~8-9GB when threads = 4
+
+            # pick lower value of jobs based on cores vs memory metric to minimize oom and swap usage during compilation
+            max_jobs = max(1, min(max_num_jobs_cores, max_num_jobs_memory))
+            os.environ["MAX_JOBS"] = str(max_jobs)
+
+        super().__init__(*args, **kwargs)
+
+
+setup(
+    name=PACKAGE_NAME,
+    version=get_package_version(),
+    packages=find_packages(
+        exclude=(
+            "build",
+            "csrc",
+            "include",
+            "tests",
+            "dist",
+            "docs",
+            "benchmarks",
+            "flash_attn.egg-info",
+        )
+    ),
+    author="Tri Dao",
+    author_email="[email protected]",
+    description="Flash Attention: Fast and Memory-Efficient Exact Attention",
+    long_description=long_description,
+    long_description_content_type="text/markdown",
+    url="https://github.com/Dao-AILab/flash-attention",
+    classifiers=[
+        "Programming Language :: Python :: 3",
+        "License :: OSI Approved :: BSD License",
+        "Operating System :: Unix",
+    ],
+    ext_modules=ext_modules,
+    cmdclass={"bdist_wheel": CachedWheelsCommand, "build_ext": NinjaBuildExtension}
+    if ext_modules
+    else {
+        "bdist_wheel": CachedWheelsCommand,
+    },
+    python_requires=">=3.8",
+    install_requires=[
+        "torch",
+        "einops",
+    ],
+    setup_requires=[
+        "packaging",
+        "psutil",
+        "ninja",
+    ],
+)

flash-attention/tests/models/test_vit.py

@@ -0,0 +1,48 @@
+import re
+
+import pytest
+import torch
+from flash_attn.models.vit import vit_base_patch16_224 as flash_vit_base_patch16_224
+from timm.models.vision_transformer import vit_base_patch16_224
+
+
+@pytest.mark.parametrize("fused_mlp", [False, True])
+# @pytest.mark.parametrize('fused_mlp', [False])
+@pytest.mark.parametrize("optimized", [False, True])
+# @pytest.mark.parametrize('optimized', [True])
+def test_vit(optimized, fused_mlp):
+    """Check that our implementation of ViT matches the timm's implementation:
+    the output of our forward pass in fp16 should be around the same as
+    timm' forward pass in fp16, when compared to timm's forward pass in fp32.
+    """
+    dtype = torch.float16
+    device = "cuda"
+
+    kwargs = {}
+    if optimized:
+        kwargs = dict(use_flash_attn=True, fused_bias_fc=True, fused_dropout_add_ln=True)
+    kwargs["fused_mlp"] = fused_mlp
+    model = flash_vit_base_patch16_224(**kwargs).to(device=device, dtype=dtype)
+
+    model_ref = vit_base_patch16_224(pretrained=True).to(device=device)
+    model_timm = vit_base_patch16_224(pretrained=True).to(device=device, dtype=dtype)
+
+    model.load_state_dict(model_ref.state_dict())
+
+    model.eval()
+    model_ref.eval()
+    model_timm.eval()
+
+    torch.manual_seed(0)
+    batch_size = 2
+    x = torch.randn(batch_size, 3, 224, 224, device=device, dtype=dtype)
+    out = model(x)
+    out_timm = model_timm(x)
+    out_ref = model_ref(x.float())
+
+    print(f"Output max diff: {(out - out_ref).abs().max().item()}")
+    print(f"Output mean diff: {(out - out_ref).abs().mean().item()}")
+    print(f"timm fp16 max diff: {(out_timm - out_ref).abs().max().item()}")
+    print(f"timm fp16 mean diff: {(out_timm - out_ref).abs().mean().item()}")
+    rtol = 2 if not fused_mlp else 8
+    assert (out - out_ref).abs().max().item() < rtol * (out_timm - out_ref).abs().max().item()

flash-attention/tests/modules/test_embedding_parallel.py

@@ -0,0 +1,106 @@
+# Run test with:
+# torchrun --no_python --nproc_per_node=8 pytest -q -s tests/modules/test_embedding_parallel.py
+
+import pytest
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from apex.transformer import parallel_state
+from einops import rearrange
+from flash_attn.modules.embedding import GPT2Embeddings, ParallelGPT2Embeddings
+
+is_sm8x = torch.cuda.get_device_capability("cuda")[0] >= 8
+
+
+@pytest.mark.parametrize("dtype", [torch.float16] + ([torch.bfloat16] if is_sm8x else []))
+# @pytest.mark.parametrize('dtype', [torch.bfloat16])
+@pytest.mark.parametrize("world_size", [1, 2, 4, 8])
+# @pytest.mark.parametrize('world_size', [2])
+@pytest.mark.parametrize("sequence_parallel", [True, False])
+# @pytest.mark.parametrize('sequence_parallel', [False])
+@pytest.mark.parametrize("has_pos_emb", [True, False])
+# @pytest.mark.parametrize('has_pos_emb', [True])
+@pytest.mark.parametrize("dim", [1024])
+def test_embedding_parallel(dim, has_pos_emb, sequence_parallel, world_size, dtype):
+    vocab_size = 50264
+    seqlen = 2048
+    assert vocab_size % world_size == 0
+    assert dim % world_size == 0
+    rtol, atol = (3e-3, 5e-2) if dtype == torch.bfloat16 else (3e-3, 3e-3)
+    if not torch.distributed.is_initialized():
+        torch.distributed.init_process_group(backend="nccl", init_method="env://")
+    device = f"cuda:{torch.distributed.get_rank()}"
+    assert world_size <= torch.distributed.get_world_size()
+    parallel_state.initialize_model_parallel(tensor_model_parallel_size_=world_size)
+    rank = parallel_state.get_tensor_model_parallel_rank()
+    # set seed
+    torch.random.manual_seed(0)
+    batch_size = 8
+    seqlen = 1024
+    assert (batch_size * seqlen) % world_size == 0
+    input_ids_pt = torch.randint(0, vocab_size, (batch_size, seqlen), device=device)
+    input_ids = input_ids_pt.detach().clone()
+
+    model_pt = GPT2Embeddings(
+        dim, vocab_size, seqlen if has_pos_emb else 0, device=device, dtype=dtype
+    )
+    model = ParallelGPT2Embeddings(
+        dim,
+        vocab_size,
+        seqlen if has_pos_emb else 0,
+        parallel_state.get_tensor_model_parallel_group(),
+        sequence_parallel=sequence_parallel,
+        device=device,
+        dtype=dtype,
+    )
+    partition_vocab_size = vocab_size // world_size
+    partition_dim = dim // world_size
+    with torch.no_grad():
+        model.word_embeddings.weight.copy_(
+            model_pt.word_embeddings.weight[
+                rank * partition_vocab_size : (rank + 1) * partition_vocab_size
+            ]
+        )
+        if has_pos_emb:
+            model.position_embeddings.weight.copy_(
+                model_pt.position_embeddings.weight[
+                    :, rank * partition_dim : (rank + 1) * partition_dim
+                ]
+            )
+
+    out = model(input_ids, combine_batch_seqlen_dim=True)
+    out_pt = rearrange(model_pt(input_ids), "b s d -> (b s) d")
+    partition_batch_dim = batch_size * seqlen // world_size
+    assert torch.allclose(
+        out,
+        out_pt[rank * partition_batch_dim : (rank + 1) * partition_batch_dim]
+        if sequence_parallel
+        else out_pt,
+        rtol=rtol,
+        atol=atol,
+    )
+
+    g = torch.randn_like(out_pt)
+    out_pt.backward(g)
+    out.backward(
+        g[rank * partition_batch_dim : (rank + 1) * partition_batch_dim] if sequence_parallel else g
+    )
+    parallel_state.destroy_model_parallel()
+
+    assert torch.allclose(
+        model.word_embeddings.weight.grad,
+        model_pt.word_embeddings.weight.grad[
+            rank * partition_vocab_size : (rank + 1) * partition_vocab_size
+        ],
+        rtol=rtol,
+        atol=atol,
+    )
+    if has_pos_emb:
+        assert torch.allclose(
+            model.position_embeddings.weight.grad,
+            model_pt.position_embeddings.weight.grad[
+                :, rank * partition_dim : (rank + 1) * partition_dim
+            ],
+            rtol=rtol,
+            atol=atol,
+        )

flash-attention/tests/modules/test_mha_parallel.py

@@ -0,0 +1,160 @@
+# Run test with:
+# torchrun --no_python --nproc_per_node=8 pytest -q -s tests/modules/test_mha_parallel.py
+
+import math
+
+import pytest
+import torch
+import torch.nn.functional as F
+from apex.transformer import parallel_state, tensor_parallel
+from einops import rearrange
+from flash_attn.modules.mha import MHA, ParallelMHA
+
+is_sm8x = torch.cuda.get_device_capability("cuda")[0] >= 8
+
+
+@pytest.mark.parametrize("dtype", [torch.float16] + ([torch.bfloat16] if is_sm8x else []))
+# @pytest.mark.parametrize('dtype', [torch.float16])
+@pytest.mark.parametrize("world_size", [1, 2, 4, 8])
+# @pytest.mark.parametrize('world_size', [2])
+@pytest.mark.parametrize("sequence_parallel", [True, False])
+# @pytest.mark.parametrize('sequence_parallel', [False])
+@pytest.mark.parametrize("head_dim", [64, 128])
+# @pytest.mark.parametrize('head_dim', [64])
+@pytest.mark.parametrize("embed_dim", [1024, 4096])
+# @pytest.mark.parametrize('embed_dim', [1024])
+def test_mha_parallel(embed_dim, head_dim, sequence_parallel, world_size, dtype):
+    assert embed_dim % head_dim == 0
+    num_heads = embed_dim // head_dim
+    assert num_heads % world_size == 0
+    rtol, atol = (3e-3, 1e-2) if dtype == torch.bfloat16 else (3e-3, 1e-3)
+    if not torch.distributed.is_initialized():
+        torch.distributed.init_process_group(backend="nccl", init_method="env://")
+    device = f"cuda:{torch.distributed.get_rank()}"
+    assert world_size <= torch.distributed.get_world_size()
+    parallel_state.initialize_model_parallel(tensor_model_parallel_size_=world_size)
+    rank = parallel_state.get_tensor_model_parallel_rank()
+    # set seed
+    torch.random.manual_seed(0)
+    batch_size = 2
+    seqlen = 1024
+    assert (batch_size * seqlen) % world_size == 0
+    x_pt = torch.randn(
+        batch_size * seqlen, embed_dim, device=device, dtype=dtype, requires_grad=True
+    )
+    # We need to generate g here so that all processes get the same gradient,
+    # as rank 0 will have an extra bias that changes the RNG.
+    # If we don't divide by batch_size, the gradient gets a bit too large.
+    g = torch.randn_like(x_pt) / 32
+    if sequence_parallel:
+        x = (
+            tensor_parallel.scatter_to_sequence_parallel_region(x_pt)
+            .detach()
+            .clone()
+            .requires_grad_()
+        )
+    else:
+        x = x_pt.detach().clone().requires_grad_()
+
+    model_pt = MHA(
+        embed_dim,
+        num_heads,
+        rotary_emb_dim=int(head_dim // 2),
+        use_flash_attn=True,
+        device=device,
+        dtype=dtype,
+    )
+    partition_dim = embed_dim // world_size
+    model = ParallelMHA(
+        embed_dim,
+        num_heads,
+        parallel_state.get_tensor_model_parallel_group(),
+        rotary_emb_dim=int(head_dim // 2),
+        use_flash_attn=True,
+        sequence_parallel=sequence_parallel,
+        device=device,
+        dtype=dtype,
+    )
+
+    with torch.no_grad():
+        model.Wqkv.weight.copy_(
+            rearrange(
+                rearrange(model_pt.Wqkv.weight, "(three o) i -> three o i", three=3)[
+                    :, rank * partition_dim : (rank + 1) * partition_dim
+                ],
+                "three o i -> (three o) i",
+            )
+        )
+        model.Wqkv.bias.copy_(
+            rearrange(
+                rearrange(model_pt.Wqkv.bias, "(three o) -> three o", three=3)[
+                    :, rank * partition_dim : (rank + 1) * partition_dim
+                ],
+                "three o -> (three o)",
+            )
+        )
+        model.out_proj.weight.copy_(
+            model_pt.out_proj.weight[:, rank * partition_dim : (rank + 1) * partition_dim]
+        )
+        if rank == 0:
+            model.out_proj.bias.copy_(model_pt.out_proj.bias)
+
+    out = model(x, seqlen=seqlen)
+    out_pt = rearrange(model_pt(rearrange(x_pt, "(b s) d -> b s d", s=seqlen)), "b s d -> (b s) d")
+    partition_batch_dim = batch_size * seqlen // world_size
+    assert torch.allclose(
+        out,
+        out_pt[rank * partition_batch_dim : (rank + 1) * partition_batch_dim]
+        if sequence_parallel
+        else out_pt,
+        rtol=rtol,
+        atol=atol,
+    )
+
+    out_pt.backward(g)
+    out.backward(
+        g[rank * partition_batch_dim : (rank + 1) * partition_batch_dim] if sequence_parallel else g
+    )
+    parallel_state.destroy_model_parallel()
+
+    assert torch.allclose(
+        x.grad,
+        x_pt.grad[rank * partition_batch_dim : (rank + 1) * partition_batch_dim]
+        if sequence_parallel
+        else x_pt.grad,
+        rtol=rtol,
+        atol=atol / 100,  # magnitude of x.grad is quite small
+    )
+    # The error for d_weight and d_bias is quite a bit higher
+    assert torch.allclose(
+        model.Wqkv.weight.grad,
+        rearrange(
+            rearrange(model_pt.Wqkv.weight.grad, "(three o) i -> three o i", three=3)[
+                :, rank * partition_dim : (rank + 1) * partition_dim
+            ],
+            "three o i -> (three o) i",
+        ),
+        rtol=rtol,
+        atol=atol * 10,
+    )
+    assert torch.allclose(
+        model.Wqkv.bias.grad,
+        rearrange(
+            rearrange(model_pt.Wqkv.bias.grad, "(three o) -> three o", three=3)[
+                :, rank * partition_dim : (rank + 1) * partition_dim
+            ],
+            "three o -> (three o)",
+        ),
+        rtol=rtol,
+        atol=atol * 5,
+    )
+    assert torch.allclose(
+        model.out_proj.weight.grad,
+        model_pt.out_proj.weight.grad[:, rank * partition_dim : (rank + 1) * partition_dim],
+        rtol=rtol,
+        atol=atol * 10,
+    )
+    if rank == 0:
+        assert torch.allclose(
+            model.out_proj.bias.grad, model_pt.out_proj.bias.grad, rtol=rtol, atol=atol * 5
+        )

flash-attention/tests/modules/test_mlp_parallel.py

@@ -0,0 +1,143 @@
+# Run test with:
+# torchrun --no_python --nproc_per_node=8 pytest -q -s tests/modules/test_mlp_parallel.py
+
+import pytest
+import torch
+import torch.nn.functional as F
+from apex.transformer import parallel_state, tensor_parallel
+from einops import rearrange
+from flash_attn.modules.mlp import GatedMlp, ParallelGatedMlp
+
+is_sm8x = torch.cuda.get_device_capability("cuda")[0] >= 8
+
+
+@pytest.mark.parametrize("dtype", [torch.float16] + ([torch.bfloat16] if is_sm8x else []))
+# @pytest.mark.parametrize('dtype', [torch.float16])
+@pytest.mark.parametrize("world_size", [1, 2, 4, 8])
+# @pytest.mark.parametrize('world_size', [2])
+@pytest.mark.parametrize("sequence_parallel", [True, False])
+# @pytest.mark.parametrize('sequence_parallel', [False])
+@pytest.mark.parametrize("activation", [F.silu, F.sigmoid])
+# @pytest.mark.parametrize('activation', [F.silu])
+@pytest.mark.parametrize("dim", [1024, 4096])
+# @pytest.mark.parametrize('dim', [1024])
+def test_mlp_parallel(dim, activation, sequence_parallel, world_size, dtype):
+    rtol, atol = (3e-3, 3e-2) if dtype == torch.bfloat16 else (3e-3, 3e-3)
+
+    if not torch.distributed.is_initialized():
+        torch.distributed.init_process_group(backend="nccl", init_method="env://")
+    device = f"cuda:{torch.distributed.get_rank()}"
+    assert world_size <= torch.distributed.get_world_size()
+    parallel_state.initialize_model_parallel(tensor_model_parallel_size_=world_size)
+    rank = parallel_state.get_tensor_model_parallel_rank()
+    # set seed
+    torch.random.manual_seed(0)
+    batch_size = 2
+    seqlen = 1024
+    assert (batch_size * seqlen) % world_size == 0
+    x_pt = torch.randn(batch_size * seqlen, dim, device=device, dtype=dtype, requires_grad=True)
+    # We need to generate g here so that all processes get the same gradient,
+    # as rank 0 will have an extra bias that changes the RNG.
+    # If we don't divide by batch_size, the gradient gets a bit too large.
+    g = torch.randn_like(x_pt) / 32
+    if sequence_parallel:
+        x = (
+            tensor_parallel.scatter_to_sequence_parallel_region(x_pt)
+            .detach()
+            .clone()
+            .requires_grad_()
+        )
+    else:
+        x = x_pt.detach().clone().requires_grad_()
+
+    model_pt = GatedMlp(dim, activation=activation, device=device, dtype=dtype)
+    partition_dim = model_pt.fc1.weight.shape[0] // 2 // world_size
+    model = ParallelGatedMlp(
+        dim,
+        parallel_state.get_tensor_model_parallel_group(),
+        activation=activation,
+        sequence_parallel=sequence_parallel,
+        device=device,
+        dtype=dtype,
+    )
+
+    with torch.no_grad():
+        model.fc1.weight.copy_(
+            rearrange(
+                rearrange(model_pt.fc1.weight, "(two o) i -> two o i", two=2)[
+                    :, rank * partition_dim : (rank + 1) * partition_dim
+                ],
+                "two o i -> (two o) i",
+            )
+        )
+        model.fc1.bias.copy_(
+            rearrange(
+                rearrange(model_pt.fc1.bias, "(two o) -> two o", two=2)[
+                    :, rank * partition_dim : (rank + 1) * partition_dim
+                ],
+                "two o -> (two o)",
+            )
+        )
+        model.fc2.weight.copy_(
+            model_pt.fc2.weight[:, rank * partition_dim : (rank + 1) * partition_dim]
+        )
+        if rank == 0:
+            model.fc2.bias.copy_(model_pt.fc2.bias)
+
+    out = model(x)
+    out_pt = model_pt(x_pt)
+    partition_batch_dim = batch_size * seqlen // world_size
+    assert torch.allclose(
+        out,
+        out_pt[rank * partition_batch_dim : (rank + 1) * partition_batch_dim]
+        if sequence_parallel
+        else out_pt,
+        rtol=rtol,
+        atol=atol,
+    )
+
+    out_pt.backward(g)
+    out.backward(
+        g[rank * partition_batch_dim : (rank + 1) * partition_batch_dim] if sequence_parallel else g
+    )
+    parallel_state.destroy_model_parallel()
+
+    assert torch.allclose(
+        x.grad,
+        x_pt.grad[rank * partition_batch_dim : (rank + 1) * partition_batch_dim]
+        if sequence_parallel
+        else x_pt.grad,
+        rtol=rtol,
+        atol=atol,
+    )
+
+    assert torch.allclose(
+        model.fc1.weight.grad,
+        rearrange(
+            rearrange(model_pt.fc1.weight.grad, "(two o) i -> two o i", two=2)[
+                :, rank * partition_dim : (rank + 1) * partition_dim
+            ],
+            "two o i -> (two o) i",
+        ),
+        rtol=rtol,
+        atol=atol,
+    )
+    assert torch.allclose(
+        model.fc1.bias.grad,
+        rearrange(
+            rearrange(model_pt.fc1.bias.grad, "(two o) -> two o", two=2)[
+                :, rank * partition_dim : (rank + 1) * partition_dim
+            ],
+            "two o -> (two o)",
+        ),
+        rtol=rtol,
+        atol=atol,
+    )
+    assert torch.allclose(
+        model.fc2.weight.grad,
+        model_pt.fc2.weight.grad[:, rank * partition_dim : (rank + 1) * partition_dim],
+        rtol=rtol,
+        atol=atol,
+    )
+    if rank == 0:
+        assert torch.allclose(model.fc2.bias.grad, model_pt.fc2.bias.grad, rtol=rtol, atol=atol)

flash-attention/tests/ops/test_dropout_layer_norm.py

@@ -0,0 +1,1189 @@
+import math
+
+import pytest
+import torch
+import torch.nn.functional as F
+from einops import rearrange, repeat
+from flash_attn.ops.layer_norm import (
+    DropoutAddLayerNorm,
+    dropout_add_layer_norm,
+    dropout_add_layer_norm_parallel_residual,
+    dropout_add_layer_norm_subset,
+)
+from flash_attn.ops.rms_norm import (
+    DropoutAddRMSNorm,
+    dropout_add_rms_norm,
+    dropout_add_rms_norm_parallel_residual,
+    dropout_add_rms_norm_subset,
+)
+
+try:
+    from apex.normalization import FusedRMSNorm
+    from apex.normalization.fused_layer_norm import fused_rms_norm_affine
+except:
+    FusedRMSNorm, fused_rms_norm_affine = None, None
+
+
+is_sm8x = torch.cuda.get_device_capability("cuda")[0] >= 8
+
+
+@pytest.mark.parametrize("is_rms_norm", [False, True])
+@pytest.mark.parametrize("has_colscale", [True, False])
+# @pytest.mark.parametrize('has_colscale', [False])
+@pytest.mark.parametrize("has_rowscale", [True, False])
+# @pytest.mark.parametrize('has_rowscale', [True])
+@pytest.mark.parametrize("has_residual", [True, False])
+# @pytest.mark.parametrize('has_residual', [False])
+@pytest.mark.parametrize("dropout_p", [0.37, 0.0])
+# @pytest.mark.parametrize('dropout_p', [0.0])
+@pytest.mark.parametrize("weight_dtype", [torch.float32, torch.float16])
+# @pytest.mark.parametrize('weight_dtype', [torch.float32])
+@pytest.mark.parametrize(
+    "input_dtype,residual_dtype",
+    [(torch.float16, torch.float16), (torch.float16, torch.float32), (torch.float32, torch.float32)]
+    + ([(torch.bfloat16, torch.bfloat16), (torch.bfloat16, torch.float32)] if is_sm8x else []),
+)
+# @pytest.mark.parametrize('input_dtype,residual_dtype', [(torch.float16, torch.float32)])
+@pytest.mark.parametrize(
+    "hidden_size",
+    [192, 256, 384, 768, 1024, 1280, 1536, 1600, 2048, 2560, 3000, 3072, 4096, 5120, 6144],
+)
+# @pytest.mark.parametrize('hidden_size', [256])
+def test_dropout_layer_norm_training(
+    hidden_size,
+    input_dtype,
+    residual_dtype,
+    weight_dtype,
+    dropout_p,
+    has_residual,
+    has_rowscale,
+    has_colscale,
+    is_rms_norm,
+):
+    if weight_dtype == torch.float16 and input_dtype == torch.bfloat16:
+        pytest.skip()  # Not supported
+    if is_rms_norm and FusedRMSNorm is None:
+        pytest.skip()  # We need Apex's FusedRMSNorm to test
+    layer_norm_cls = torch.nn.LayerNorm if not is_rms_norm else FusedRMSNorm
+    our_layer_norm_cls = DropoutAddLayerNorm if not is_rms_norm else DropoutAddRMSNorm
+    our_layer_norm_func = dropout_add_layer_norm if not is_rms_norm else dropout_add_rms_norm
+    device = "cuda"
+    # rtol, atol = (1e-5, 1e-6) if input_dtype == torch.float32 else (1e-3, 1e-4)
+    rtol, atol = (1e-3, 1e-4)
+    # set seed
+    torch.random.manual_seed(0)
+    batch_size = 8
+    seqlen = 512
+    x0_pt = torch.randn(
+        batch_size, seqlen, hidden_size, device=device, dtype=input_dtype, requires_grad=True
+    )
+    x0 = x0_pt.detach().clone().requires_grad_()
+    x0_ref = x0_pt.detach().clone().float().requires_grad_()
+    if has_colscale:
+        colscale = torch.randn(hidden_size, device=device, dtype=weight_dtype, requires_grad=True)
+        colscale_pt = colscale.detach().clone().requires_grad_()
+        colscale_ref = colscale.detach().clone().float().requires_grad_()
+    else:
+        colscale = None
+    if has_residual:
+        res_pt = torch.randn_like(x0, dtype=residual_dtype, requires_grad=True)
+        res = res_pt.detach().clone().requires_grad_()
+        res_ref = res_pt.detach().clone().float().requires_grad_()
+    else:
+        res = None
+    if has_rowscale:
+        rowscale = torch.empty(batch_size, seqlen, device=device, dtype=input_dtype)
+        survival_rate = 0.87
+        rowscale = rowscale.bernoulli_(survival_rate) / survival_rate
+        x0_scaled_pt = x0_pt * rearrange(rowscale, "... -> ... 1")
+        x0_scaled_ref = x0_ref * rearrange(rowscale, "... -> ... 1")
+    else:
+        rowscale = None
+        x0_scaled_pt = x0_pt
+        x0_scaled_ref = x0_ref
+    if has_colscale:
+        x0_scaled_pt = x0_scaled_pt * colscale_pt
+        x0_scaled_ref = x0_scaled_ref * colscale_ref
+    model_pt = layer_norm_cls(hidden_size).to(device=device, dtype=weight_dtype)
+    torch.nn.init.normal_(model_pt.weight)
+    if not is_rms_norm:
+        torch.nn.init.normal_(model_pt.bias)
+    model_ref = layer_norm_cls(hidden_size).to(device=device, dtype=torch.float32)
+    model = our_layer_norm_cls(hidden_size, p=dropout_p, device=device, dtype=weight_dtype)
+    with torch.no_grad():
+        model.weight.copy_(model_pt.weight)
+        model_ref.weight.copy_(model_pt.weight)
+        if not is_rms_norm:
+            model.bias.copy_(model_pt.bias)
+            model_ref.bias.copy_(model_pt.bias)
+    residual_in_fp32 = (not has_residual) and residual_dtype == torch.float32
+    out, dmask = our_layer_norm_func(
+        x0,
+        res,
+        model.weight,
+        model.bias,
+        model.p,
+        model.eps,
+        rowscale=rowscale,
+        layerscale=colscale,
+        residual_in_fp32=residual_in_fp32,
+        return_dropout_mask=True,
+    )
+    assert out.dtype == input_dtype
+    print(f"Actual dropout fraction: {1 - dmask.float().mean().item()}")
+    if has_residual:
+        residual_pt = (
+            (x0_scaled_pt.float() * dmask.float()) / (1 - dropout_p) + res_pt.float()
+        ).to(dtype=residual_dtype)
+        residual_ref = (x0_scaled_ref * dmask.float()) / (1 - dropout_p) + res_ref
+    else:
+        residual_pt = ((x0_scaled_pt.float() * dmask.float()) / (1 - dropout_p)).to(
+            dtype=residual_dtype
+        )
+        residual_ref = (x0_scaled_ref * dmask.float()) / (1 - dropout_p)
+    out_pt = model_pt(residual_pt.to(dtype=weight_dtype)).to(dtype=input_dtype)
+    out_ref = model_ref(residual_ref)
+    assert (out - out_ref).abs().max() <= 4 * (out_pt - out_ref).abs().max() + 1e-4
+
+    g = torch.randn_like(out) / batch_size
+    out_pt.backward(g)
+    out.backward(g)
+    out_ref.backward(g)
+    assert (x0.grad - x0_ref.grad).abs().max() <= 4 * (x0_pt.grad - x0_ref.grad).abs().max() + 1e-4
+    if has_residual:
+        assert (res.grad - res_ref.grad).abs().max() <= 4 * (
+            res_pt.grad - res_ref.grad
+        ).abs().max() + 1e-4
+    assert (model.weight.grad - model_ref.weight.grad).abs().max() <= 3 * (
+        model_pt.weight.grad - model_ref.weight.grad
+    ).abs().max() + 3e-5
+    if not is_rms_norm:
+        assert (model.bias.grad - model_ref.bias.grad).abs().max() <= 2 * (
+            model_pt.bias.grad - model_ref.bias.grad
+        ).abs().max() + 3e-5
+    if has_colscale:
+        assert (colscale.grad - colscale_ref.grad).abs().max() <= 2 * (
+            colscale_pt.grad - colscale_ref.grad
+        ).abs().max() + 2e-4
+
+
+@pytest.mark.parametrize("weight_dtype", [torch.float32, torch.float16])
+@pytest.mark.parametrize(
+    "input_dtype,residual_dtype",
+    [(torch.float16, torch.float16), (torch.float16, torch.float32), (torch.float32, torch.float32)]
+    + ([(torch.bfloat16, torch.bfloat16), (torch.bfloat16, torch.float32)] if is_sm8x else []),
+)
+@pytest.mark.parametrize("hidden_size", [768, 1024, 1280, 1536, 1600, 2048, 2560, 3072, 4096, 5120])
+def test_dropout_layer_norm_eval(hidden_size, input_dtype, residual_dtype, weight_dtype):
+    if weight_dtype == torch.float16 and input_dtype == torch.bfloat16:
+        pytest.skip()  # Not supported
+    device = "cuda"
+    # rtol, atol = (1e-5, 1e-6) if dtype == torch.float32 else (1e-3, 1e-4)
+    rtol, atol = (1e-3, 1e-4)
+    dropout_p = 0.37
+    # set seed
+    torch.random.manual_seed(0)
+    batch_size = 32
+    seqlen = 512
+    x0_pt = torch.randn(
+        batch_size, seqlen, hidden_size, device=device, dtype=input_dtype, requires_grad=True
+    )
+    x0 = x0_pt.detach().clone().requires_grad_()
+    x0_ref = x0_pt.detach().clone().float().requires_grad_()
+    res_pt = torch.randn_like(x0, dtype=residual_dtype, requires_grad=True)
+    res = res_pt.detach().clone().requires_grad_()
+    res_ref = res_pt.detach().clone().float().requires_grad_()
+    model_pt = torch.nn.LayerNorm(hidden_size, device=device, dtype=weight_dtype)
+    torch.nn.init.normal_(model_pt.weight)
+    torch.nn.init.normal_(model_pt.bias)
+    model = DropoutAddLayerNorm(hidden_size, p=dropout_p, device=device, dtype=weight_dtype)
+    model_ref = torch.nn.LayerNorm(hidden_size, device=device, dtype=torch.float32)
+    with torch.no_grad():
+        model.weight.copy_(model_pt.weight)
+        model.bias.copy_(model_pt.bias)
+        model_ref.weight.copy_(model_pt.weight)
+        model_ref.bias.copy_(model_pt.bias)
+    model_pt.eval()
+    model.eval()
+    model_ref.eval()
+    out = model(x0, res)
+    residual_pt = (x0_pt.float() + res_pt.float()).to(dtype=residual_dtype)
+    residual_ref = x0_ref + res_ref
+    out_pt = model_pt(residual_pt.to(dtype=weight_dtype)).to(input_dtype)
+    out_ref = model_ref(residual_ref)
+    assert (out - out_ref).abs().max() <= 4 * (out_pt - out_ref).abs().max() + 1e-4
+
+
+@pytest.mark.parametrize("is_rms_norm", [False, True])
+@pytest.mark.parametrize("has_colscale", [True, False])
+@pytest.mark.parametrize("has_rowscale", [True, False])
+@pytest.mark.parametrize("has_residual", [True, False])
+@pytest.mark.parametrize("dropout_p", [0.37, 0.0])
+@pytest.mark.parametrize("weight_dtype", [torch.float32, torch.float16])
+@pytest.mark.parametrize(
+    "input_dtype,residual_dtype",
+    [(torch.float16, torch.float16), (torch.float16, torch.float32), (torch.float32, torch.float32)]
+    + ([(torch.bfloat16, torch.bfloat16), (torch.bfloat16, torch.float32)] if is_sm8x else []),
+)
+# @pytest.mark.parametrize('has_colscale', [True])
+# @pytest.mark.parametrize('has_rowscale', [False])
+# @pytest.mark.parametrize('has_residual', [True])
+# @pytest.mark.parametrize('dropout_p', [0.0])
+# @pytest.mark.parametrize('weight_dtype', [torch.float32])
+# @pytest.mark.parametrize('input_dtype,residual_dtype', [(torch.float32, torch.float32)])
+@pytest.mark.parametrize(
+    "hidden_size",
+    [192, 256, 384, 768, 1024, 1280, 1536, 1600, 2048, 2560, 3000, 3072, 4096, 5120, 6144],
+)
+# @pytest.mark.parametrize('hidden_size', [256])
+def test_dropout_layer_norm_prenorm_training(
+    hidden_size,
+    input_dtype,
+    residual_dtype,
+    weight_dtype,
+    dropout_p,
+    has_residual,
+    has_rowscale,
+    has_colscale,
+    is_rms_norm,
+):
+    if weight_dtype == torch.float16 and input_dtype == torch.bfloat16:
+        pytest.skip()  # Not supported
+    if is_rms_norm and FusedRMSNorm is None:
+        pytest.skip()  # We need Apex's FusedRMSNorm to test
+    layer_norm_cls = torch.nn.LayerNorm if not is_rms_norm else FusedRMSNorm
+    our_layer_norm_cls = DropoutAddLayerNorm if not is_rms_norm else DropoutAddRMSNorm
+    our_layer_norm_func = dropout_add_layer_norm if not is_rms_norm else dropout_add_rms_norm
+    device = "cuda"
+    # rtol, atol = (1e-5, 1e-6) if input_dtype == torch.float32 else (1e-3, 1e-4)
+    rtol, atol = (1e-3, 2e-4)
+    # set seed
+    torch.random.manual_seed(0)
+    batch_size = 8
+    seqlen = 512
+    x0_pt = torch.randn(
+        batch_size, seqlen, hidden_size, device=device, dtype=input_dtype, requires_grad=True
+    )
+    x0 = x0_pt.detach().clone().requires_grad_()
+    x0_ref = x0_pt.detach().clone().float().requires_grad_()
+    if has_colscale:
+        colscale = torch.randn(hidden_size, device=device, dtype=weight_dtype, requires_grad=True)
+        colscale_pt = colscale.detach().clone().requires_grad_()
+        colscale_ref = colscale.detach().clone().float().requires_grad_()
+    else:
+        colscale = None
+    if has_residual:
+        res_pt = torch.randn_like(x0, dtype=residual_dtype, requires_grad=True)
+        res = res_pt.detach().clone().requires_grad_()
+        res_ref = res_pt.detach().clone().float().requires_grad_()
+    else:
+        res = None
+    if has_rowscale:
+        rowscale = torch.empty(batch_size, seqlen, device=device, dtype=input_dtype)
+        survival_rate = 0.87
+        rowscale = rowscale.bernoulli_(survival_rate) / survival_rate
+        x0_scaled_pt = x0_pt * rearrange(rowscale, "... -> ... 1")
+        x0_scaled_ref = x0_ref * rearrange(rowscale, "... -> ... 1")
+    else:
+        rowscale = None
+        x0_scaled_pt = x0_pt
+        x0_scaled_ref = x0_ref
+    if has_colscale:
+        x0_scaled_pt = x0_scaled_pt * colscale_pt
+        x0_scaled_ref = x0_scaled_ref * colscale_ref
+    model_pt = layer_norm_cls(hidden_size).to(device=device, dtype=weight_dtype)
+    torch.nn.init.normal_(model_pt.weight)
+    if not is_rms_norm:
+        torch.nn.init.normal_(model_pt.bias)
+    model_ref = layer_norm_cls(hidden_size).to(device=device, dtype=torch.float32)
+    model = our_layer_norm_cls(
+        hidden_size, prenorm=True, p=dropout_p, device=device, dtype=weight_dtype
+    )
+    with torch.no_grad():
+        model.weight.copy_(model_pt.weight)
+        model_ref.weight.copy_(model_pt.weight)
+        if not is_rms_norm:
+            model.bias.copy_(model_pt.bias)
+            model_ref.bias.copy_(model_pt.bias)
+    residual_in_fp32 = (not has_residual) and residual_dtype == torch.float32
+    out, residual, dmask = our_layer_norm_func(
+        x0,
+        res,
+        model.weight,
+        model.bias,
+        model.p,
+        model.eps,
+        rowscale=rowscale,
+        layerscale=colscale,
+        prenorm=True,
+        residual_in_fp32=residual_in_fp32,
+        return_dropout_mask=True,
+    )
+    print(f"Actual dropout fraction: {1 - dmask.float().mean().item()}")
+    if has_residual:
+        residual_pt = (
+            (x0_scaled_pt.float() * dmask.float()) / (1 - dropout_p) + res_pt.float()
+        ).to(dtype=residual_dtype)
+        residual_ref = (x0_scaled_ref * dmask.float()) / (1 - dropout_p) + res_ref
+    else:
+        residual_pt = ((x0_scaled_pt.float() * dmask.float()) / (1 - dropout_p)).to(
+            dtype=residual_dtype
+        )
+        residual_ref = (x0_scaled_ref * dmask.float()) / (1 - dropout_p)
+    out_pt = model_pt(residual_pt.to(dtype=weight_dtype)).to(dtype=input_dtype)
+    out_ref = model_ref(residual_ref)
+    assert out.dtype == input_dtype
+    assert residual.dtype == residual_dtype
+    assert (out - out_ref).abs().max() <= 4 * (out_pt - out_ref).abs().max() + 1e-4
+    assert (residual - residual_ref).abs().max() <= 4 * (
+        residual_pt - residual_ref
+    ).abs().max() + 1e-4
+
+    g = torch.randn_like(out) / batch_size
+    (out_pt * F.sigmoid(residual_pt)).backward(g)
+    (out * F.sigmoid(residual)).backward(g)
+    (out_ref * F.sigmoid(residual_ref.to(dtype=residual_dtype))).backward(g)
+    assert (x0.grad - x0_ref.grad).abs().max() <= 4 * (x0_pt.grad - x0_ref.grad).abs().max() + 1e-4
+    if has_residual:
+        assert (res.grad - res_ref.grad).abs().max() <= 4 * (
+            res_pt.grad - res_ref.grad
+        ).abs().max() + 1e-4
+    assert (model.weight.grad - model_ref.weight.grad).abs().max() <= 2 * (
+        model_pt.weight.grad - model_ref.weight.grad
+    ).abs().max() + 2e-4
+    if not is_rms_norm:
+        assert (model.bias.grad - model_ref.bias.grad).abs().max() <= 2 * (
+            model_pt.bias.grad - model_ref.bias.grad
+        ).abs().max() + 2e-4
+    if has_colscale:
+        assert (colscale.grad - colscale_ref.grad).abs().max() <= 2 * (
+            colscale_pt.grad - colscale_ref.grad
+        ).abs().max() + 2e-4
+
+
+@pytest.mark.parametrize("weight_dtype", [torch.float32, torch.float16])
+@pytest.mark.parametrize(
+    "input_dtype,residual_dtype",
+    [(torch.float16, torch.float16), (torch.float16, torch.float32), (torch.float32, torch.float32)]
+    + ([(torch.bfloat16, torch.bfloat16), (torch.bfloat16, torch.float32)] if is_sm8x else []),
+)
+@pytest.mark.parametrize("hidden_size", [768, 1024, 1280, 1536, 1600, 2048, 2560, 3072, 4096, 5120])
+def test_dropout_layer_norm_prenorm_eval(hidden_size, input_dtype, residual_dtype, weight_dtype):
+    if weight_dtype == torch.float16 and input_dtype == torch.bfloat16:
+        pytest.skip()  # Not supported
+    device = "cuda"
+    # rtol, atol = (1e-5, 1e-6) if dtype == torch.float32 else (1e-3, 1e-4)
+    rtol, atol = (1e-3, 1e-4)
+    dropout_p = 0.37
+    # set seed
+    torch.random.manual_seed(0)
+    batch_size = 32
+    seqlen = 512
+    x0_pt = torch.randn(
+        batch_size, seqlen, hidden_size, device=device, dtype=input_dtype, requires_grad=True
+    )
+    x0 = x0_pt.detach().clone().requires_grad_()
+    x0_ref = x0_pt.detach().clone().float().requires_grad_()
+    res_pt = torch.randn_like(x0, dtype=residual_dtype, requires_grad=True)
+    res = res_pt.detach().clone().requires_grad_()
+    res_ref = res_pt.detach().clone().float().requires_grad_()
+    model_pt = torch.nn.LayerNorm(hidden_size, device=device, dtype=weight_dtype)
+    torch.nn.init.normal_(model_pt.weight)
+    torch.nn.init.normal_(model_pt.bias)
+    model = DropoutAddLayerNorm(
+        hidden_size, prenorm=True, p=dropout_p, device=device, dtype=weight_dtype
+    )
+    model_ref = torch.nn.LayerNorm(hidden_size, device=device, dtype=torch.float32)
+    with torch.no_grad():
+        model.weight.copy_(model_pt.weight)
+        model.bias.copy_(model_pt.bias)
+        model_ref.weight.copy_(model_pt.weight)
+        model_ref.bias.copy_(model_pt.bias)
+    model_pt.eval()
+    model.eval()
+    model_ref.eval()
+    out, residual = model(x0, res)
+    residual_pt = (x0_pt.float() + res_pt.float()).to(dtype=residual_dtype)
+    residual_ref = x0_ref + res_ref
+    out_pt = model_pt(residual_pt.to(dtype=weight_dtype)).to(input_dtype)
+    out_ref = model_ref(residual_ref)
+    assert (out - out_ref).abs().max() <= 4 * (out_pt - out_ref).abs().max() + 1e-4
+    assert (residual - residual_ref).abs().max() <= 4 * (
+        residual_pt - residual_ref
+    ).abs().max() + 1e-4
+
+
+@pytest.mark.parametrize("has_colscale", [True, False])
+@pytest.mark.parametrize("has_residual", [True, False])
+@pytest.mark.parametrize("dropout_p", [0.37, 0.0])
+@pytest.mark.parametrize("weight_dtype", [torch.float32, torch.float16])
+@pytest.mark.parametrize(
+    "input_dtype,residual_dtype",
+    [(torch.float16, torch.float16), (torch.float16, torch.float32), (torch.float32, torch.float32)]
+    + ([(torch.bfloat16, torch.bfloat16), (torch.bfloat16, torch.float32)] if is_sm8x else []),
+)
+# @pytest.mark.parametrize('has_colscale', [True])
+# @pytest.mark.parametrize('has_residual', [True])
+# @pytest.mark.parametrize('dropout_p', [0.0])
+# @pytest.mark.parametrize('weight_dtype', [torch.float32])
+# @pytest.mark.parametrize('input_dtype,residual_dtype', [(torch.float32, torch.float32)])
+@pytest.mark.parametrize(
+    "hidden_size",
+    [192, 256, 384, 768, 1024, 1280, 1536, 1600, 2048, 2560, 3000, 3072, 4096, 5120, 6144],
+)
+# @pytest.mark.parametrize('hidden_size', [256])
+def test_dropout_layer_norm_subset_training(
+    hidden_size, input_dtype, residual_dtype, weight_dtype, dropout_p, has_residual, has_colscale
+):
+    if weight_dtype == torch.float16 and input_dtype == torch.bfloat16:
+        pytest.skip()  # Not supported
+    device = "cuda"
+    # rtol, atol = (1e-5, 1e-6) if input_dtype == torch.float32 else (1e-3, 1e-4)
+    rtol, atol = (1e-3, 2e-4)
+    # set seed
+    torch.random.manual_seed(0)
+    batch_size = 8
+    seqlen = 512
+    drop_path_rate = 0.4
+    drop_path_scale = 1 / (1 - drop_path_rate)
+
+    def generate_droppath_masks(batch_size, seqlen, drop_path_rate, device):
+        # Do it on CPU so we can get the numrows (with .item()) without GPU-CPU sync
+        mask_batch = torch.rand(batch_size) < 1 - drop_path_rate
+        numrows = (mask_batch).sum().item() * seqlen
+        mask_batch = mask_batch.to(device=device, non_blocking=True)
+        mask_batch_seqlen = repeat(mask_batch, "b -> (b s)", s=seqlen)
+        subset = torch.cumsum(mask_batch_seqlen, dim=0, dtype=torch.int32).masked_fill_(
+            ~mask_batch_seqlen, 0
+        )
+        return mask_batch, numrows, rearrange(subset, "(b s) -> b s", b=batch_size)
+
+    x0_mask_batch, x0_numrows, x0_subset = generate_droppath_masks(
+        batch_size, seqlen, drop_path_rate, device
+    )
+    out_mask_batch, out_numrows, out_subset = generate_droppath_masks(
+        batch_size, seqlen, drop_path_rate, device
+    )
+
+    x0_pt = torch.randn(
+        batch_size, seqlen, hidden_size, device=device, dtype=input_dtype, requires_grad=True
+    )
+    x0 = x0_pt.detach().clone()[x0_mask_batch].requires_grad_()
+    x0_ref = x0_pt.detach().clone().float().requires_grad_()
+    if has_colscale:
+        colscale = torch.randn(hidden_size, device=device, dtype=weight_dtype, requires_grad=True)
+        colscale_pt = colscale.detach().clone().requires_grad_()
+        colscale_ref = colscale.detach().clone().float().requires_grad_()
+    else:
+        colscale = None
+    if has_residual:
+        res_pt = torch.randn_like(x0_pt, dtype=residual_dtype, requires_grad=True)
+        res = res_pt.detach().clone().requires_grad_()
+        res_ref = res_pt.detach().clone().float().requires_grad_()
+    else:
+        res = None
+
+    if has_colscale:
+        x0_scaled_pt = x0_pt * colscale_pt
+        x0_scaled_ref = x0_ref * colscale_ref
+    else:
+        x0_scaled_pt = x0_pt
+        x0_scaled_ref = x0_ref
+
+    model_pt = torch.nn.LayerNorm(hidden_size, device=device, dtype=weight_dtype)
+    torch.nn.init.normal_(model_pt.weight)
+    torch.nn.init.normal_(model_pt.bias)
+    model_ref = torch.nn.LayerNorm(hidden_size, device=device, dtype=torch.float32)
+    model = DropoutAddLayerNorm(
+        hidden_size, prenorm=False, p=dropout_p, device=device, dtype=weight_dtype
+    )
+    with torch.no_grad():
+        model.weight.copy_(model_pt.weight)
+        model.bias.copy_(model_pt.bias)
+        model_ref.weight.copy_(model_pt.weight)
+        model_ref.bias.copy_(model_pt.bias)
+
+    residual_in_fp32 = (not has_residual) and residual_dtype == torch.float32
+    out, dmask = dropout_add_layer_norm_subset(
+        x0,
+        res,
+        model.weight,
+        model.bias,
+        model.p,
+        model.eps,
+        layerscale=colscale,
+        x0_subset=x0_subset,
+        out_subset=out_subset,
+        rowscale_const=drop_path_scale,
+        out_numrows=out_numrows,
+        prenorm=False,
+        residual_in_fp32=residual_in_fp32,
+        return_dropout_mask=True,
+    )
+    print(f"Actual dropout fraction: {1 - dmask.float().mean().item()}")
+
+    x0_scaled_pt = (
+        x0_scaled_pt.masked_fill(repeat(~x0_mask_batch, "b -> b s d", s=seqlen, d=hidden_size), 0)
+        * drop_path_scale
+    )
+    x0_scaled_ref = (
+        x0_scaled_ref.masked_fill(repeat(~x0_mask_batch, "b -> b s d", s=seqlen, d=hidden_size), 0)
+        * drop_path_scale
+    )
+    dmask_expanded = torch.zeros_like(x0_pt, dtype=torch.uint8)
+    dmask_expanded[x0_mask_batch] = dmask
+    if has_residual:
+        residual_pt = (
+            (x0_scaled_pt.float() * dmask_expanded.float()) / (1 - dropout_p) + res_pt.float()
+        ).to(dtype=residual_dtype)
+        residual_ref = (x0_scaled_ref * dmask_expanded.float()) / (1 - dropout_p) + res_ref
+    else:
+        residual_pt = ((x0_scaled_pt.float() * dmask_expanded.float()) / (1 - dropout_p)).to(
+            dtype=residual_dtype
+        )
+        residual_ref = (x0_scaled_ref * dmask_expanded.float()) / (1 - dropout_p)
+    out_pt = model_pt(residual_pt.to(dtype=weight_dtype)).to(dtype=input_dtype)[out_mask_batch]
+    out_ref = model_ref(residual_ref)[out_mask_batch]
+    assert out.dtype == input_dtype
+    assert (out - out_ref).abs().max() <= 4 * (out_pt - out_ref).abs().max() + 1e-4
+
+    g = torch.randn_like(out) / batch_size
+    out_pt.backward(g)
+    out.backward(g)
+    out_ref.backward(g)
+    assert (x0.grad - x0_ref.grad[x0_mask_batch]).abs().max() <= 4 * (x0_pt.grad - x0_ref.grad)[
+        x0_mask_batch
+    ].abs().max() + 1e-4
+    if has_residual:
+        assert (res.grad - res_ref.grad).abs().max() <= 4 * (
+            res_pt.grad - res_ref.grad
+        ).abs().max() + 1e-4
+    assert (model.weight.grad - model_ref.weight.grad).abs().max() <= 2 * (
+        model_pt.weight.grad - model_ref.weight.grad
+    ).abs().max() + 2e-4
+    assert (model.bias.grad - model_ref.bias.grad).abs().max() <= 2 * (
+        model_pt.bias.grad - model_ref.bias.grad
+    ).abs().max() + 2e-4
+    if has_colscale:
+        assert (colscale.grad - colscale_ref.grad).abs().max() <= 2 * (
+            colscale_pt.grad - colscale_ref.grad
+        ).abs().max() + 2e-4
+
+
+@pytest.mark.parametrize("has_colscale", [True, False])
+@pytest.mark.parametrize("has_residual", [True, False])
+@pytest.mark.parametrize("dropout_p", [0.37, 0.0])
+@pytest.mark.parametrize("weight_dtype", [torch.float32, torch.float16])
+@pytest.mark.parametrize(
+    "input_dtype,residual_dtype",
+    [(torch.float16, torch.float16), (torch.float16, torch.float32), (torch.float32, torch.float32)]
+    + ([(torch.bfloat16, torch.bfloat16), (torch.bfloat16, torch.float32)] if is_sm8x else []),
+)
+# @pytest.mark.parametrize('has_colscale', [True])
+# @pytest.mark.parametrize('has_residual', [True])
+# @pytest.mark.parametrize('dropout_p', [0.0])
+# @pytest.mark.parametrize('weight_dtype', [torch.float32])
+# @pytest.mark.parametrize('input_dtype,residual_dtype', [(torch.float32, torch.float32)])
+@pytest.mark.parametrize(
+    "hidden_size",
+    [192, 256, 384, 768, 1024, 1280, 1536, 1600, 2048, 2560, 3000, 3072, 4096, 5120, 6144],
+)
+# @pytest.mark.parametrize('hidden_size', [256])
+def test_dropout_layer_norm_subset_prenorm_training(
+    hidden_size, input_dtype, residual_dtype, weight_dtype, dropout_p, has_residual, has_colscale
+):
+    if weight_dtype == torch.float16 and input_dtype == torch.bfloat16:
+        pytest.skip()  # Not supported
+    device = "cuda"
+    # rtol, atol = (1e-5, 1e-6) if input_dtype == torch.float32 else (1e-3, 1e-4)
+    rtol, atol = (1e-3, 2e-4)
+    # set seed
+    torch.random.manual_seed(0)
+    batch_size = 8
+    seqlen = 512
+    drop_path_rate = 0.4
+    drop_path_scale = 1 / (1 - drop_path_rate)
+
+    def generate_droppath_masks(batch_size, seqlen, drop_path_rate, device):
+        # Do it on CPU so we can get the numrows (with .item()) without GPU-CPU sync
+        mask_batch = torch.rand(batch_size) < 1 - drop_path_rate
+        numrows = (mask_batch).sum().item() * seqlen
+        mask_batch = mask_batch.to(device=device, non_blocking=True)
+        mask_batch_seqlen = repeat(mask_batch, "b -> (b s)", s=seqlen)
+        subset = torch.cumsum(mask_batch_seqlen, dim=0, dtype=torch.int32).masked_fill_(
+            ~mask_batch_seqlen, 0
+        )
+        return mask_batch, numrows, rearrange(subset, "(b s) -> b s", b=batch_size)
+
+    x0_mask_batch, x0_numrows, x0_subset = generate_droppath_masks(
+        batch_size, seqlen, drop_path_rate, device
+    )
+    out_mask_batch, out_numrows, out_subset = generate_droppath_masks(
+        batch_size, seqlen, drop_path_rate, device
+    )
+
+    x0_pt = torch.randn(
+        batch_size, seqlen, hidden_size, device=device, dtype=input_dtype, requires_grad=True
+    )
+    x0 = x0_pt.detach().clone()[x0_mask_batch].requires_grad_()
+    x0_ref = x0_pt.detach().clone().float().requires_grad_()
+    if has_colscale:
+        colscale = torch.randn(hidden_size, device=device, dtype=weight_dtype, requires_grad=True)
+        colscale_pt = colscale.detach().clone().requires_grad_()
+        colscale_ref = colscale.detach().clone().float().requires_grad_()
+    else:
+        colscale = None
+    if has_residual:
+        res_pt = torch.randn_like(x0_pt, dtype=residual_dtype, requires_grad=True)
+        res = res_pt.detach().clone().requires_grad_()
+        res_ref = res_pt.detach().clone().float().requires_grad_()
+    else:
+        res = None
+
+    if has_colscale:
+        x0_scaled_pt = x0_pt * colscale_pt
+        x0_scaled_ref = x0_ref * colscale_ref
+    else:
+        x0_scaled_pt = x0_pt
+        x0_scaled_ref = x0_ref
+
+    model_pt = torch.nn.LayerNorm(hidden_size, device=device, dtype=weight_dtype)
+    torch.nn.init.normal_(model_pt.weight)
+    torch.nn.init.normal_(model_pt.bias)
+    model_ref = torch.nn.LayerNorm(hidden_size, device=device, dtype=torch.float32)
+    model = DropoutAddLayerNorm(
+        hidden_size, prenorm=True, p=dropout_p, device=device, dtype=weight_dtype
+    )
+    with torch.no_grad():
+        model.weight.copy_(model_pt.weight)
+        model.bias.copy_(model_pt.bias)
+        model_ref.weight.copy_(model_pt.weight)
+        model_ref.bias.copy_(model_pt.bias)
+
+    residual_in_fp32 = (not has_residual) and residual_dtype == torch.float32
+    out, residual, dmask = dropout_add_layer_norm_subset(
+        x0,
+        res,
+        model.weight,
+        model.bias,
+        model.p,
+        model.eps,
+        layerscale=colscale,
+        x0_subset=x0_subset,
+        out_subset=out_subset,
+        rowscale_const=drop_path_scale,
+        out_numrows=out_numrows,
+        prenorm=True,
+        residual_in_fp32=residual_in_fp32,
+        return_dropout_mask=True,
+    )
+    print(f"Actual dropout fraction: {1 - dmask.float().mean().item()}")
+
+    x0_scaled_pt = (
+        x0_scaled_pt.masked_fill(repeat(~x0_mask_batch, "b -> b s d", s=seqlen, d=hidden_size), 0)
+        * drop_path_scale
+    )
+    x0_scaled_ref = (
+        x0_scaled_ref.masked_fill(repeat(~x0_mask_batch, "b -> b s d", s=seqlen, d=hidden_size), 0)
+        * drop_path_scale
+    )
+    dmask_expanded = torch.zeros_like(x0_pt, dtype=torch.uint8)
+    dmask_expanded[x0_mask_batch] = dmask
+    if has_residual:
+        residual_pt = (
+            (x0_scaled_pt.float() * dmask_expanded.float()) / (1 - dropout_p) + res_pt.float()
+        ).to(dtype=residual_dtype)
+        residual_ref = (x0_scaled_ref * dmask_expanded.float()) / (1 - dropout_p) + res_ref
+    else:
+        residual_pt = ((x0_scaled_pt.float() * dmask_expanded.float()) / (1 - dropout_p)).to(
+            dtype=residual_dtype
+        )
+        residual_ref = (x0_scaled_ref * dmask_expanded.float()) / (1 - dropout_p)
+    out_pt = model_pt(residual_pt.to(dtype=weight_dtype)).to(dtype=input_dtype)[out_mask_batch]
+    out_ref = model_ref(residual_ref)[out_mask_batch]
+    assert out.dtype == input_dtype
+    assert residual.dtype == residual_dtype
+    assert (out - out_ref).abs().max() <= 4 * (out_pt - out_ref).abs().max() + 1e-4
+    assert (residual - residual_ref).abs().max() <= 4 * (
+        residual_pt - residual_ref
+    ).abs().max() + 1e-4
+
+    g = torch.randn_like(out) / batch_size
+    (out_pt * F.sigmoid(residual_pt[out_mask_batch]) + residual_pt.mean(0, keepdim=True)).backward(
+        g
+    )
+    (out * F.sigmoid(residual[out_mask_batch]) + residual.mean(0, keepdim=True)).backward(g)
+    (
+        out_ref * F.sigmoid(residual_ref[out_mask_batch].to(dtype=residual_dtype))
+        + residual_ref.mean(0, keepdim=True)
+    ).backward(g)
+    assert (x0.grad - x0_ref.grad[x0_mask_batch]).abs().max() <= 4 * (x0_pt.grad - x0_ref.grad)[
+        x0_mask_batch
+    ].abs().max() + 1e-4
+    if has_residual:
+        assert (res.grad - res_ref.grad).abs().max() <= 4 * (
+            res_pt.grad - res_ref.grad
+        ).abs().max() + 1e-4
+    assert (model.weight.grad - model_ref.weight.grad).abs().max() <= 2 * (
+        model_pt.weight.grad - model_ref.weight.grad
+    ).abs().max() + 2e-4
+    assert (model.bias.grad - model_ref.bias.grad).abs().max() <= 2 * (
+        model_pt.bias.grad - model_ref.bias.grad
+    ).abs().max() + 2e-4
+    if has_colscale:
+        assert (colscale.grad - colscale_ref.grad).abs().max() <= 2 * (
+            colscale_pt.grad - colscale_ref.grad
+        ).abs().max() + 2e-4
+
+
+@pytest.mark.parametrize("is_rms_norm", [False, True])
+# @pytest.mark.parametrize('is_rms_norm', [False])
+@pytest.mark.parametrize("tied_norm", [False, True])
+# @pytest.mark.parametrize('tied_norm', [False])
+@pytest.mark.parametrize("has_residual", [True, False])
+# @pytest.mark.parametrize('has_residual', [False])
+@pytest.mark.parametrize("has_x1", [True, False])
+# @pytest.mark.parametrize('has_x1', [True])
+@pytest.mark.parametrize("dropout_p", [0.37, 0.0])
+# @pytest.mark.parametrize('dropout_p', [0.0])
+@pytest.mark.parametrize("weight_dtype", [torch.float32, torch.float16])
+# @pytest.mark.parametrize('weight_dtype', [torch.float16])
+@pytest.mark.parametrize(
+    "input_dtype,residual_dtype",
+    [(torch.float16, torch.float16), (torch.float16, torch.float32), (torch.float32, torch.float32)]
+    + ([(torch.bfloat16, torch.bfloat16), (torch.bfloat16, torch.float32)] if is_sm8x else []),
+)
+# @pytest.mark.parametrize('input_dtype,residual_dtype', [(torch.float16, torch.float32)])
+@pytest.mark.parametrize(
+    "hidden_size",
+    [192, 256, 384, 768, 1024, 1280, 1536, 1600, 2048, 2560, 3000, 3072, 4096, 5120, 6144],
+)
+# @pytest.mark.parametrize('hidden_size', [256])
+def test_dropout_layer_norm_parallel_residual_training(
+    hidden_size,
+    input_dtype,
+    residual_dtype,
+    weight_dtype,
+    dropout_p,
+    has_x1,
+    has_residual,
+    tied_norm,
+    is_rms_norm,
+):
+    if weight_dtype == torch.float16 and input_dtype == torch.bfloat16:
+        pytest.skip()  # Not supported
+    if is_rms_norm and fused_rms_norm_affine is None:
+        pytest.skip()  # We need Apex's FusedRMSNorm to test
+    our_layer_norm_func = (
+        dropout_add_layer_norm_parallel_residual
+        if not is_rms_norm
+        else dropout_add_rms_norm_parallel_residual
+    )
+    device = "cuda"
+    # rtol, atol = (1e-5, 1e-6) if input_dtype == torch.float32 else (1e-3, 1e-4)
+    rtol, atol = (1e-3, 1e-4)
+    # set seed
+    torch.random.manual_seed(0)
+    batch_size = 8
+    seqlen = 512
+    x0_pt = torch.randn(
+        batch_size, seqlen, hidden_size, device=device, dtype=input_dtype, requires_grad=True
+    )
+    x0 = x0_pt.detach().clone().requires_grad_()
+    x0_ref = x0_pt.detach().clone().float().requires_grad_()
+    if has_x1:
+        x1_pt = torch.randn(
+            batch_size, seqlen, hidden_size, device=device, dtype=input_dtype, requires_grad=True
+        )
+        x1 = x1_pt.detach().clone().requires_grad_()
+        x1_ref = x1_pt.detach().clone().float().requires_grad_()
+    else:
+        x1 = None
+    if has_residual:
+        res_pt = torch.randn_like(x0, dtype=residual_dtype, requires_grad=True)
+        res = res_pt.detach().clone().requires_grad_()
+        res_ref = res_pt.detach().clone().float().requires_grad_()
+    else:
+        res = None
+    weight0 = torch.randn(hidden_size, device=device, dtype=weight_dtype, requires_grad=True)
+    bias0 = (
+        torch.randn(hidden_size, device=device, dtype=weight_dtype, requires_grad=True)
+        if not is_rms_norm
+        else None
+    )
+    weight0_pt = weight0.detach().clone().requires_grad_()
+    weight0_ref = weight0.detach().clone().float().requires_grad_()
+    bias0_pt = bias0.detach().clone().requires_grad_() if bias0 is not None else None
+    bias0_ref = bias0.detach().clone().float().requires_grad_() if bias0 is not None else None
+    if not tied_norm:
+        weight1 = torch.randn(hidden_size, device=device, dtype=weight_dtype, requires_grad=True)
+        bias1 = (
+            torch.randn(hidden_size, device=device, dtype=weight_dtype, requires_grad=True)
+            if not is_rms_norm
+            else None
+        )
+        weight1_pt = weight1.detach().clone().requires_grad_()
+        weight1_ref = weight1.detach().clone().float().requires_grad_()
+        bias1_pt = bias1.detach().clone().requires_grad_() if bias1 is not None else None
+        bias1_ref = bias1.detach().clone().float().requires_grad_() if bias1 is not None else None
+    else:
+        weight1, bias1 = None, None
+    epsilon = 1e-5
+    residual_in_fp32 = (not has_residual) and residual_dtype == torch.float32
+
+    out0, out1, dmask0, dmask1 = our_layer_norm_func(
+        x0,
+        x1,
+        res,
+        weight0,
+        bias0,
+        weight1,
+        bias1,
+        dropout_p,
+        epsilon,
+        residual_in_fp32=residual_in_fp32,
+        return_dropout_mask=True,
+    )
+    assert out0.dtype == input_dtype
+    if not tied_norm:
+        assert out1.dtype == input_dtype
+    print(f"Actual dropout fraction: {1 - dmask0.float().mean().item()}")
+    if has_residual:
+        if has_x1:
+            residual_pt = (
+                (x0_pt.float() * dmask0.float()) / (1 - dropout_p)
+                + (x1_pt.float() * dmask1.float()) / (1 - dropout_p)
+                + res_pt.float()
+            ).to(dtype=residual_dtype)
+            residual_ref = (
+                (x0_ref * dmask0.float()) / (1 - dropout_p)
+                + (x1_ref * dmask1.float()) / (1 - dropout_p)
+            ) + res_ref
+        else:
+            residual_pt = ((x0_pt.float() * dmask0.float()) / (1 - dropout_p) + res_pt.float()).to(
+                dtype=residual_dtype
+            )
+            residual_ref = (x0_ref * dmask0.float()) / (1 - dropout_p) + res_ref
+    else:
+        if has_x1:
+            residual_pt = (
+                (x0_pt.float() * dmask0.float()) / (1 - dropout_p)
+                + (x1_pt.float() * dmask1.float()) / (1 - dropout_p)
+            ).to(dtype=residual_dtype)
+            residual_ref = (x0_ref * dmask0.float()) / (1 - dropout_p) + (
+                x1_ref * dmask1.float()
+            ) / (1 - dropout_p)
+        else:
+            residual_pt = ((x0_pt.float() * dmask0.float()) / (1 - dropout_p)).to(
+                dtype=residual_dtype
+            )
+            residual_ref = (x0_ref * dmask0.float()) / (1 - dropout_p)
+    if not is_rms_norm:
+        out0_pt = F.layer_norm(
+            residual_pt.to(dtype=weight_dtype), (hidden_size,), weight0_pt, bias0_pt, eps=epsilon
+        ).to(dtype=input_dtype)
+        out0_ref = F.layer_norm(residual_ref, (hidden_size,), weight0_ref, bias0_ref, eps=epsilon)
+        if not tied_norm:
+            out1_pt = F.layer_norm(
+                residual_pt.to(dtype=weight_dtype),
+                (hidden_size,),
+                weight1_pt,
+                bias1_pt,
+                eps=epsilon,
+            ).to(dtype=input_dtype)
+            out1_ref = F.layer_norm(
+                residual_ref, (hidden_size,), weight1_ref, bias1_ref, eps=epsilon
+            )
+    else:
+        out0_pt = fused_rms_norm_affine(
+            residual_pt.to(dtype=weight_dtype), weight0_pt, (hidden_size,), eps=epsilon
+        ).to(dtype=input_dtype)
+        out0_ref = fused_rms_norm_affine(residual_ref, weight0_ref, (hidden_size,), eps=epsilon)
+        if not tied_norm:
+            out1_pt = fused_rms_norm_affine(
+                residual_pt.to(dtype=weight_dtype), weight1_pt, (hidden_size,), eps=epsilon
+            ).to(dtype=input_dtype)
+            out1_ref = fused_rms_norm_affine(residual_ref, weight1_ref, (hidden_size,), eps=epsilon)
+
+    assert (out0 - out0_ref).abs().max() <= 4 * (out0_pt - out0_ref).abs().max() + 1e-4
+    if not tied_norm:
+        assert (out1 - out1_ref).abs().max() <= 4 * (out1_pt - out1_ref).abs().max() + 1e-4
+
+    g0 = torch.randn_like(out0) / batch_size
+    if tied_norm:
+        out0.backward(g0)
+        out0_pt.backward(g0)
+        out0_ref.backward(g0)
+    else:
+        g1 = torch.randn_like(out1) / batch_size
+        (out0 * g0 + out1 * g1).sum().backward()
+        (out0_pt * g0 + out1_pt * g1).sum().backward()
+        (out0_ref * g0 + out1_ref * g1).sum().backward()
+    assert (x0.grad - x0_ref.grad).abs().max() <= 4 * (x0_pt.grad - x0_ref.grad).abs().max() + 1e-4
+    if has_x1:
+        assert (x1.grad - x1_ref.grad).abs().max() <= 4 * (
+            x1_pt.grad - x1_ref.grad
+        ).abs().max() + 1e-4
+    if has_residual:
+        assert (res.grad - res_ref.grad).abs().max() <= 4 * (
+            res_pt.grad - res_ref.grad
+        ).abs().max() + 1e-4
+    assert (weight0.grad - weight0_ref.grad).abs().max() <= 3 * (
+        weight0_pt.grad - weight0_ref.grad
+    ).abs().max() + 3e-5
+    if not is_rms_norm:
+        assert (bias0.grad - bias0_ref.grad).abs().max() <= 2 * (
+            bias0_pt.grad - bias0_ref.grad
+        ).abs().max() + 3e-5
+    if not tied_norm:
+        assert (weight1.grad - weight1_ref.grad).abs().max() <= 3 * (
+            weight1_pt.grad - weight1_ref.grad
+        ).abs().max() + 3e-5
+        if not is_rms_norm:
+            assert (bias1.grad - bias1_ref.grad).abs().max() <= 2 * (
+                bias1_pt.grad - bias1_ref.grad
+            ).abs().max() + 3e-5
+
+
+@pytest.mark.parametrize("is_rms_norm", [False, True])
+# @pytest.mark.parametrize('is_rms_norm', [False])
+@pytest.mark.parametrize("tied_norm", [False, True])
+# @pytest.mark.parametrize('tied_norm', [False])
+@pytest.mark.parametrize("has_residual", [True, False])
+# @pytest.mark.parametrize('has_residual', [False])
+@pytest.mark.parametrize("has_x1", [True, False])
+# @pytest.mark.parametrize('has_x1', [True])
+@pytest.mark.parametrize("dropout_p", [0.37, 0.0])
+# @pytest.mark.parametrize('dropout_p', [0.0])
+@pytest.mark.parametrize("weight_dtype", [torch.float32, torch.float16])
+# @pytest.mark.parametrize('weight_dtype', [torch.float16])
+@pytest.mark.parametrize(
+    "input_dtype,residual_dtype",
+    [(torch.float16, torch.float16), (torch.float16, torch.float32), (torch.float32, torch.float32)]
+    + ([(torch.bfloat16, torch.bfloat16), (torch.bfloat16, torch.float32)] if is_sm8x else []),
+)
+# @pytest.mark.parametrize('input_dtype,residual_dtype', [(torch.float16, torch.float32)])
+@pytest.mark.parametrize(
+    "hidden_size",
+    [192, 256, 384, 768, 1024, 1280, 1536, 1600, 2048, 2560, 3000, 3072, 4096, 5120, 6144],
+)
+# @pytest.mark.parametrize('hidden_size', [256])
+def test_dropout_layer_norm_parallel_residual_prenorm_training(
+    hidden_size,
+    input_dtype,
+    residual_dtype,
+    weight_dtype,
+    dropout_p,
+    has_x1,
+    has_residual,
+    tied_norm,
+    is_rms_norm,
+):
+    if weight_dtype == torch.float16 and input_dtype == torch.bfloat16:
+        pytest.skip()  # Not supported
+    if is_rms_norm and fused_rms_norm_affine is None:
+        pytest.skip()  # We need Apex's FusedRMSNorm to test
+    our_layer_norm_func = (
+        dropout_add_layer_norm_parallel_residual
+        if not is_rms_norm
+        else dropout_add_rms_norm_parallel_residual
+    )
+    device = "cuda"
+    # rtol, atol = (1e-5, 1e-6) if input_dtype == torch.float32 else (1e-3, 1e-4)
+    rtol, atol = (1e-3, 1e-4)
+    # set seed
+    torch.random.manual_seed(0)
+    batch_size = 8
+    seqlen = 512
+    x0_pt = torch.randn(
+        batch_size, seqlen, hidden_size, device=device, dtype=input_dtype, requires_grad=True
+    )
+    x0 = x0_pt.detach().clone().requires_grad_()
+    x0_ref = x0_pt.detach().clone().float().requires_grad_()
+    if has_x1:
+        x1_pt = torch.randn(
+            batch_size, seqlen, hidden_size, device=device, dtype=input_dtype, requires_grad=True
+        )
+        x1 = x1_pt.detach().clone().requires_grad_()
+        x1_ref = x1_pt.detach().clone().float().requires_grad_()
+    else:
+        x1 = None
+    if has_residual:
+        res_pt = torch.randn_like(x0, dtype=residual_dtype, requires_grad=True)
+        res = res_pt.detach().clone().requires_grad_()
+        res_ref = res_pt.detach().clone().float().requires_grad_()
+    else:
+        res = None
+    weight0 = torch.randn(hidden_size, device=device, dtype=weight_dtype, requires_grad=True)
+    bias0 = (
+        torch.randn(hidden_size, device=device, dtype=weight_dtype, requires_grad=True)
+        if not is_rms_norm
+        else None
+    )
+    weight0_pt = weight0.detach().clone().requires_grad_()
+    weight0_ref = weight0.detach().clone().float().requires_grad_()
+    bias0_pt = bias0.detach().clone().requires_grad_() if bias0 is not None else None
+    bias0_ref = bias0.detach().clone().float().requires_grad_() if bias0 is not None else None
+    if not tied_norm:
+        weight1 = torch.randn(hidden_size, device=device, dtype=weight_dtype, requires_grad=True)
+        bias1 = (
+            torch.randn(hidden_size, device=device, dtype=weight_dtype, requires_grad=True)
+            if not is_rms_norm
+            else None
+        )
+        weight1_pt = weight1.detach().clone().requires_grad_()
+        weight1_ref = weight1.detach().clone().float().requires_grad_()
+        bias1_pt = bias1.detach().clone().requires_grad_() if bias1 is not None else None
+        bias1_ref = bias1.detach().clone().float().requires_grad_() if bias1 is not None else None
+    else:
+        weight1, bias1 = None, None
+    epsilon = 1e-5
+    residual_in_fp32 = (not has_residual) and residual_dtype == torch.float32
+
+    out0, out1, residual, dmask0, dmask1 = our_layer_norm_func(
+        x0,
+        x1,
+        res,
+        weight0,
+        bias0,
+        weight1,
+        bias1,
+        dropout_p,
+        epsilon,
+        prenorm=True,
+        residual_in_fp32=residual_in_fp32,
+        return_dropout_mask=True,
+    )
+    assert out0.dtype == input_dtype
+    if not tied_norm:
+        assert out1.dtype == input_dtype
+    print(f"Actual dropout fraction: {1 - dmask0.float().mean().item()}")
+    if has_residual:
+        if has_x1:
+            residual_pt = (
+                (x0_pt.float() * dmask0.float()) / (1 - dropout_p)
+                + (x1_pt.float() * dmask1.float()) / (1 - dropout_p)
+                + res_pt.float()
+            ).to(dtype=residual_dtype)
+            residual_ref = (
+                (x0_ref * dmask0.float()) / (1 - dropout_p)
+                + (x1_ref * dmask1.float()) / (1 - dropout_p)
+            ) + res_ref
+        else:
+            residual_pt = ((x0_pt.float() * dmask0.float()) / (1 - dropout_p) + res_pt.float()).to(
+                dtype=residual_dtype
+            )
+            residual_ref = (x0_ref * dmask0.float()) / (1 - dropout_p) + res_ref
+    else:
+        if has_x1:
+            residual_pt = (
+                (x0_pt.float() * dmask0.float()) / (1 - dropout_p)
+                + (x1_pt.float() * dmask1.float()) / (1 - dropout_p)
+            ).to(dtype=residual_dtype)
+            residual_ref = (x0_ref * dmask0.float()) / (1 - dropout_p) + (
+                x1_ref * dmask1.float()
+            ) / (1 - dropout_p)
+        else:
+            residual_pt = ((x0_pt.float() * dmask0.float()) / (1 - dropout_p)).to(
+                dtype=residual_dtype
+            )
+            residual_ref = (x0_ref * dmask0.float()) / (1 - dropout_p)
+    if not is_rms_norm:
+        out0_pt = F.layer_norm(
+            residual_pt.to(dtype=weight_dtype), (hidden_size,), weight0_pt, bias0_pt, eps=epsilon
+        ).to(dtype=input_dtype)
+        out0_ref = F.layer_norm(residual_ref, (hidden_size,), weight0_ref, bias0_ref, eps=epsilon)
+        if not tied_norm:
+            out1_pt = F.layer_norm(
+                residual_pt.to(dtype=weight_dtype),
+                (hidden_size,),
+                weight1_pt,
+                bias1_pt,
+                eps=epsilon,
+            ).to(dtype=input_dtype)
+            out1_ref = F.layer_norm(
+                residual_ref, (hidden_size,), weight1_ref, bias1_ref, eps=epsilon
+            )
+    else:
+        out0_pt = fused_rms_norm_affine(
+            residual_pt.to(dtype=weight_dtype), weight0_pt, (hidden_size,), eps=epsilon
+        ).to(dtype=input_dtype)
+        out0_ref = fused_rms_norm_affine(residual_ref, weight0_ref, (hidden_size,), eps=epsilon)
+        if not tied_norm:
+            out1_pt = fused_rms_norm_affine(
+                residual_pt.to(dtype=weight_dtype), weight1_pt, (hidden_size,), eps=epsilon
+            ).to(dtype=input_dtype)
+            out1_ref = fused_rms_norm_affine(residual_ref, weight1_ref, (hidden_size,), eps=epsilon)
+
+    assert (out0 - out0_ref).abs().max() <= 4 * (out0_pt - out0_ref).abs().max() + 1e-4
+    if not tied_norm:
+        assert (out1 - out1_ref).abs().max() <= 4 * (out1_pt - out1_ref).abs().max() + 1e-4
+    assert (residual - residual_ref).abs().max() <= 4 * (
+        residual_pt - residual_ref
+    ).abs().max() + 1e-4
+
+    g0 = torch.randn_like(out0) / batch_size
+    if tied_norm:
+        (out0 * F.sigmoid(residual)).backward(g0)
+        (out0_pt * F.sigmoid(residual_pt)).backward(g0)
+        (out0_ref * F.sigmoid(residual_ref)).backward(g0)
+    else:
+        g1 = torch.randn_like(out1) / batch_size
+        (out0 * F.sigmoid(residual) * g0 + out1 * g1).sum().backward()
+        (out0_pt * F.sigmoid(residual_pt) * g0 + out1_pt * g1).sum().backward()
+        (out0_ref * F.sigmoid(residual_ref) * g0 + out1_ref * g1).sum().backward()
+    assert (x0.grad - x0_ref.grad).abs().max() <= 4 * (x0_pt.grad - x0_ref.grad).abs().max() + 1e-4
+    if has_x1:
+        assert (x1.grad - x1_ref.grad).abs().max() <= 4 * (
+            x1_pt.grad - x1_ref.grad
+        ).abs().max() + 1e-4
+    if has_residual:
+        assert (res.grad - res_ref.grad).abs().max() <= 4 * (
+            res_pt.grad - res_ref.grad
+        ).abs().max() + 1e-4
+    assert (weight0.grad - weight0_ref.grad).abs().max() <= 3 * (
+        weight0_pt.grad - weight0_ref.grad
+    ).abs().max() + 3e-5
+    if not is_rms_norm:
+        assert (bias0.grad - bias0_ref.grad).abs().max() <= 2 * (
+            bias0_pt.grad - bias0_ref.grad
+        ).abs().max() + 3e-5
+    if not tied_norm:
+        assert (weight1.grad - weight1_ref.grad).abs().max() <= 3 * (
+            weight1_pt.grad - weight1_ref.grad
+        ).abs().max() + 3e-5
+        if not is_rms_norm:
+            assert (bias1.grad - bias1_ref.grad).abs().max() <= 2 * (
+                bias1_pt.grad - bias1_ref.grad
+            ).abs().max() + 3e-5
+
+
+def test_dropout_layer_norm_randomness():
+    hidden_size = 256
+    dtype = torch.float32
+    dropout_p = 0.1
+    device = "cuda"
+    # set seed
+    torch.random.manual_seed(0)
+    batch_size = 8
+    seqlen = 512
+    x0 = torch.randn(
+        batch_size, seqlen, hidden_size, device=device, dtype=dtype, requires_grad=True
+    )
+    res = torch.randn_like(x0, dtype=dtype, requires_grad=True)
+    model = DropoutAddLayerNorm(hidden_size, p=dropout_p, device=device, dtype=dtype)
+    torch.random.manual_seed(42)
+    _, dmask0 = dropout_add_layer_norm(
+        x0, res, model.weight, model.bias, model.p, model.eps, return_dropout_mask=True
+    )
+    # Subsequent call should have a different dropout mask
+    _, dmask1 = dropout_add_layer_norm(
+        x0, res, model.weight, model.bias, model.p, model.eps, return_dropout_mask=True
+    )
+    torch.random.manual_seed(42)
+    # Resetting the seed, should get the same dropout mask
+    _, dmask2 = dropout_add_layer_norm(
+        x0, res, model.weight, model.bias, model.p, model.eps, return_dropout_mask=True
+    )
+    assert not torch.equal(dmask0, dmask1)
+    assert torch.equal(dmask0, dmask2)

flash-attention/tests/ops/test_fused_dense.py

@@ -0,0 +1,172 @@
+import math
+from functools import partial
+
+import pytest
+import torch
+import torch.nn.functional as F
+from einops import rearrange
+from flash_attn.ops.fused_dense import FusedDense, FusedMLP
+
+
+@pytest.mark.parametrize("dtype", [torch.float16, torch.bfloat16])
+@pytest.mark.parametrize("return_residual", [False, True])
+@pytest.mark.parametrize("has_bias", [True, False])
+@pytest.mark.parametrize("out_features", [1024, 4096])
+@pytest.mark.parametrize("in_features", [1024, 4096])
+def test_fused_linear_bias(in_features, out_features, has_bias, return_residual, dtype):
+    device = "cuda"
+    rtol, atol = (3e-3, 1e-2) if dtype == torch.bfloat16 else (3e-3, 1e-3)
+    # set seed
+    torch.random.manual_seed(0)
+    batch_size = 8
+    seqlen = 512
+    x_pt = torch.randn(
+        batch_size, seqlen, in_features, device=device, dtype=dtype, requires_grad=True
+    )
+    x = x_pt.detach().clone().requires_grad_()
+    model_pt = torch.nn.Linear(in_features, out_features, bias=has_bias, device=device, dtype=dtype)
+    model = FusedDense(
+        in_features,
+        out_features,
+        bias=has_bias,
+        return_residual=return_residual,
+        device=device,
+        dtype=dtype,
+    )
+    with torch.no_grad():
+        model.weight.copy_(model_pt.weight)
+        if has_bias:
+            model.bias.copy_(model_pt.bias)
+    out_pt = model_pt(x_pt)
+    if not return_residual:
+        out = model(x)
+    else:
+        out, x_copy = model(x)
+        x_copy = (
+            x_copy[..., :out_features]
+            if out_features < in_features
+            else F.pad(x_copy, (0, out_features - in_features))
+        )
+        x_pt_copy = (
+            x_pt[..., :out_features]
+            if out_features < in_features
+            else F.pad(x_pt, (0, out_features - in_features))
+        )
+        # Just add some random function of the residual
+        out_pt = out_pt + F.gelu(x_pt_copy)
+        out = out + F.gelu(x_copy)
+
+    # with torch.no_grad():
+    #     out_fl = F.linear(x_pt.float(), model.weight.float(), model.bias.float()).half()
+    assert torch.allclose(out, out_pt, rtol=rtol, atol=atol)
+
+    # If we don't divide by batch_size, the gradient gets a bit too large.
+    g = torch.randn_like(out) / 32
+    out_pt.backward(g)
+    out.backward(g)
+    assert torch.allclose(x.grad, x_pt.grad, rtol=rtol, atol=atol)
+    # The error for d_weight and d_bias is quite a bit higher
+    assert torch.allclose(model.weight.grad, model_pt.weight.grad, rtol=rtol, atol=atol * 10)
+    if has_bias:
+        assert torch.allclose(model.bias.grad, model_pt.bias.grad, rtol=rtol, atol=atol * 5)
+
+
+@pytest.mark.parametrize("dtype", [torch.float16, torch.bfloat16])
+# @pytest.mark.parametrize('dtype', [torch.float16])
+@pytest.mark.parametrize("heuristic", ["auto", -1])
+# @pytest.mark.parametrize('heuristic', ['auto'])
+@pytest.mark.parametrize("checkpoint_lvl", [0, 1, 2])
+# @pytest.mark.parametrize('checkpoint_lvl', [1])
+@pytest.mark.parametrize("return_residual", [False, True])
+# @pytest.mark.parametrize('return_residual', [False])
+@pytest.mark.parametrize("has_bias2", [True, False])
+@pytest.mark.parametrize("has_bias1", [True, False])
+# @pytest.mark.parametrize('has_bias2', [True])
+# @pytest.mark.parametrize('has_bias1', [True])
+@pytest.mark.parametrize("activation", ["gelu_approx", "relu"])
+# @pytest.mark.parametrize('activation', ['relu'])
+@pytest.mark.parametrize("out_features", [1024, 4096])
+@pytest.mark.parametrize("in_features", [1024, 4096])
+# @pytest.mark.parametrize('out_features', [4096])
+# @pytest.mark.parametrize('in_features', [1024])
+def test_fused_mlp(
+    in_features,
+    out_features,
+    activation,
+    has_bias1,
+    has_bias2,
+    return_residual,
+    checkpoint_lvl,
+    heuristic,
+    dtype,
+):
+    device = "cuda"
+    rtol, atol = (3e-3, 3e-2) if dtype == torch.bfloat16 else (3e-3, 1e-3)
+    # set seed
+    torch.random.manual_seed(0)
+    batch_size = 8
+    seqlen = 512
+    x_pt = torch.randn(
+        batch_size, seqlen, in_features, device=device, dtype=dtype, requires_grad=True
+    )
+    x = x_pt.detach().clone().requires_grad_()
+    model_pt_fc1 = torch.nn.Linear(
+        in_features, out_features, bias=has_bias1, device=device, dtype=dtype
+    )
+    model_pt_fc2 = torch.nn.Linear(
+        out_features, in_features, bias=has_bias2, device=device, dtype=dtype
+    )
+    model = FusedMLP(
+        in_features,
+        out_features,
+        in_features,
+        activation=activation,
+        bias1=has_bias1,
+        bias2=has_bias2,
+        return_residual=return_residual,
+        checkpoint_lvl=checkpoint_lvl,
+        heuristic=heuristic,
+        device=device,
+        dtype=dtype,
+    )
+    with torch.no_grad():
+        model.fc1.weight.copy_(model_pt_fc1.weight)
+        if has_bias1:
+            model.fc1.bias.copy_(model_pt_fc1.bias)
+        model.fc2.weight.copy_(model_pt_fc2.weight)
+        if has_bias2:
+            model.fc2.bias.copy_(model_pt_fc2.bias)
+    activation_fn = (
+        partial(F.gelu, approximate="tanh")
+        if activation == "gelu_approx"
+        else partial(F.relu, inplace=True)
+    )
+    out_pt = model_pt_fc2(activation_fn(model_pt_fc1(x_pt)))
+    if not return_residual:
+        out = model(x)
+    else:
+        out, x_copy = model(x)
+        # Just add some random function of the residual
+        out_pt = out_pt + F.gelu(x_pt)
+        out = out + F.gelu(x_copy)
+    assert torch.allclose(out, out_pt, rtol=rtol, atol=atol)
+
+    # If we don't divide by batch_size, the gradient gets a bit too large.
+    g = torch.randn_like(out) / 32
+    out_pt.backward(g)
+    out.backward(g)
+    # The error for relu is higher still
+    if activation == "relu":
+        atol = 1e-1 if dtype == torch.bfloat16 else 5e-2
+    assert torch.allclose(x.grad, x_pt.grad, rtol=rtol, atol=atol)
+    # The error for d_weight and d_bias is quite a bit higher
+    assert torch.allclose(
+        model.fc1.weight.grad, model_pt_fc1.weight.grad, rtol=rtol, atol=atol * 10
+    )
+    if has_bias1:
+        assert torch.allclose(model.fc1.bias.grad, model_pt_fc1.bias.grad, rtol=rtol, atol=atol * 5)
+    assert torch.allclose(
+        model.fc2.weight.grad, model_pt_fc2.weight.grad, rtol=rtol, atol=atol * 10
+    )
+    if has_bias2:
+        assert torch.allclose(model.fc2.bias.grad, model_pt_fc2.bias.grad, rtol=rtol, atol=atol * 5)

flash-attention/tests/ops/test_fused_dense_parallel.py

@@ -0,0 +1,237 @@
+# Run test with:
+# torchrun --no_python --nproc_per_node=8 pytest -q -s tests/ops/test_fused_dense_parallel.py
+
+import math
+
+import pytest
+import torch
+import torch.nn.functional as F
+from apex.transformer import parallel_state, tensor_parallel
+from flash_attn.ops.fused_dense import ColumnParallelLinear, FusedDense, FusedMLP, ParallelFusedMLP
+
+is_sm8x = torch.cuda.get_device_capability("cuda")[0] >= 8
+
+
+@pytest.mark.parametrize("dtype", [torch.float16] + ([torch.bfloat16] if is_sm8x else []))
+# @pytest.mark.parametrize('dtype', [torch.bfloat16])
+@pytest.mark.parametrize("world_size", [1, 2, 4, 8])
+# @pytest.mark.parametrize('world_size', [2])
+@pytest.mark.parametrize("sequence_parallel", [True, False])
+# @pytest.mark.parametrize('sequence_parallel', [False])
+@pytest.mark.parametrize("has_bias", [True, False])
+# @pytest.mark.parametrize('has_bias', [False])
+@pytest.mark.parametrize("out_features", [1024])
+@pytest.mark.parametrize("in_features", [4096])
+def test_fused_linear_bias(
+    in_features, out_features, has_bias, sequence_parallel, world_size, dtype
+):
+    assert out_features % world_size == 0
+    rtol, atol = (3e-3, 3e-2) if dtype == torch.bfloat16 else (3e-3, 3e-3)
+    if not torch.distributed.is_initialized():
+        torch.distributed.init_process_group(backend="nccl", init_method="env://")
+    device = f"cuda:{torch.distributed.get_rank()}"
+    assert world_size <= torch.distributed.get_world_size()
+    parallel_state.initialize_model_parallel(tensor_model_parallel_size_=world_size)
+    rank = parallel_state.get_tensor_model_parallel_rank()
+    # set seed
+    torch.random.manual_seed(0)
+    batch_size = 2
+    seqlen = 512
+    assert batch_size * seqlen % world_size == 0
+    x_pt = torch.randn(
+        batch_size * seqlen, in_features, device=device, dtype=dtype, requires_grad=True
+    )
+    if sequence_parallel:
+        x = (
+            tensor_parallel.scatter_to_sequence_parallel_region(x_pt)
+            .detach()
+            .clone()
+            .requires_grad_()
+        )
+    else:
+        x = x_pt.detach().clone().requires_grad_()
+
+    model_pt = torch.nn.Linear(in_features, out_features, bias=has_bias, device=device, dtype=dtype)
+    partition_out_features = out_features // world_size
+    model = ColumnParallelLinear(
+        in_features,
+        out_features,
+        parallel_state.get_tensor_model_parallel_group(),
+        bias=has_bias,
+        sequence_parallel=sequence_parallel,
+        device=device,
+        dtype=dtype,
+    )
+    with torch.no_grad():
+        model.weight.copy_(
+            model_pt.weight[rank * partition_out_features : (rank + 1) * partition_out_features]
+        )
+        if has_bias:
+            model.bias.copy_(
+                model_pt.bias[rank * partition_out_features : (rank + 1) * partition_out_features]
+            )
+
+    out = model(x)
+    out_pt = model_pt(x_pt)
+    assert torch.allclose(
+        out,
+        out_pt[:, rank * partition_out_features : (rank + 1) * partition_out_features],
+        rtol=rtol,
+        atol=atol,
+    )
+
+    # If we don't divide by batch_size, the gradient gets a bit too large.
+    g = torch.randn_like(out_pt) / 32
+    out_pt.backward(g)
+    out.backward(g[:, rank * partition_out_features : (rank + 1) * partition_out_features])
+    parallel_state.destroy_model_parallel()
+
+    partition_batch_dim = batch_size * seqlen // world_size
+    assert torch.allclose(
+        x.grad,
+        x_pt.grad[rank * partition_batch_dim : (rank + 1) * partition_batch_dim]
+        if sequence_parallel
+        else x_pt.grad,
+        rtol=rtol,
+        atol=atol,
+    )
+    # The error for d_weight and d_bias is quite a bit higher
+    assert torch.allclose(
+        model.weight.grad,
+        model_pt.weight.grad[rank * partition_out_features : (rank + 1) * partition_out_features],
+        rtol=rtol,
+        atol=atol * 10,
+    )
+    if has_bias:
+        assert torch.allclose(
+            model.bias.grad,
+            model_pt.bias.grad[rank * partition_out_features : (rank + 1) * partition_out_features],
+            rtol=rtol,
+            atol=atol * 5,
+        )
+
+
+@pytest.mark.parametrize("dtype", [torch.float16] + ([torch.bfloat16] if is_sm8x else []))
+# @pytest.mark.parametrize('dtype', [torch.bfloat16])
+@pytest.mark.parametrize("world_size", [1, 2, 4, 8])
+# @pytest.mark.parametrize('world_size', [2])
+@pytest.mark.parametrize("sequence_parallel", [True, False])
+# @pytest.mark.parametrize('sequence_parallel', [False])
+@pytest.mark.parametrize("has_bias2", [True, False])
+# @pytest.mark.parametrize('has_bias2', [True])
+@pytest.mark.parametrize("out_features", [4096])
+@pytest.mark.parametrize("in_features", [1024])
+def test_fused_mlp(in_features, out_features, has_bias2, sequence_parallel, world_size, dtype):
+    assert out_features % world_size == 0
+    rtol, atol = (3e-3, 3e-2) if dtype == torch.bfloat16 else (3e-3, 3e-3)
+    if not torch.distributed.is_initialized():
+        torch.distributed.init_process_group(backend="nccl", init_method="env://")
+    device = f"cuda:{torch.distributed.get_rank()}"
+    assert world_size <= torch.distributed.get_world_size()
+    parallel_state.initialize_model_parallel(tensor_model_parallel_size_=world_size)
+    rank = parallel_state.get_tensor_model_parallel_rank()
+    # set seed
+    torch.random.manual_seed(0)
+    batch_size = 2
+    seqlen = 512
+    assert batch_size * seqlen % world_size == 0
+    x_pt = torch.randn(
+        batch_size * seqlen, in_features, device=device, dtype=dtype, requires_grad=True
+    )
+    # We need to generate g here so that all processes get the same gradient,
+    # as rank 0 will have an extra bias that changes the RNG.
+    # If we don't divide by batch_size, the gradient gets a bit too large.
+    g = torch.randn_like(x_pt) / 32
+    if sequence_parallel:
+        x = (
+            tensor_parallel.scatter_to_sequence_parallel_region(x_pt)
+            .detach()
+            .clone()
+            .requires_grad_()
+        )
+    else:
+        x = x_pt.detach().clone().requires_grad_()
+
+    model_pt_fc1 = torch.nn.Linear(in_features, out_features, device=device, dtype=dtype)
+    model_pt_fc2 = torch.nn.Linear(
+        out_features, in_features, bias=has_bias2, device=device, dtype=dtype
+    )
+    partition_out_features = out_features // world_size
+    partition_in_features = in_features // world_size
+    model = ParallelFusedMLP(
+        in_features,
+        out_features,
+        in_features,
+        process_group=parallel_state.get_tensor_model_parallel_group(),
+        bias2=has_bias2 and rank == 0,
+        sequence_parallel=sequence_parallel,
+        device=device,
+        dtype=dtype,
+    )
+
+    with torch.no_grad():
+        model.fc1.weight.copy_(
+            model_pt_fc1.weight[rank * partition_out_features : (rank + 1) * partition_out_features]
+        )
+        model.fc1.bias.copy_(
+            model_pt_fc1.bias[rank * partition_out_features : (rank + 1) * partition_out_features]
+        )
+        model.fc2.weight.copy_(
+            model_pt_fc2.weight[
+                :, rank * partition_out_features : (rank + 1) * partition_out_features
+            ]
+        )
+        if has_bias2 and rank == 0:
+            model.fc2.bias.copy_(model_pt_fc2.bias)
+
+    out = model(x)
+    out_pt = model_pt_fc2(F.gelu(model_pt_fc1(x_pt), approximate="tanh"))
+    partition_batch_dim = batch_size * seqlen // world_size
+    assert torch.allclose(
+        out,
+        out_pt[rank * partition_batch_dim : (rank + 1) * partition_batch_dim]
+        if sequence_parallel
+        else out_pt,
+        rtol=rtol,
+        atol=atol,
+    )
+
+    out_pt.backward(g)
+    out.backward(
+        g[rank * partition_batch_dim : (rank + 1) * partition_batch_dim] if sequence_parallel else g
+    )
+    parallel_state.destroy_model_parallel()
+
+    assert torch.allclose(
+        x.grad,
+        x_pt.grad[rank * partition_batch_dim : (rank + 1) * partition_batch_dim]
+        if sequence_parallel
+        else x_pt.grad,
+        rtol=rtol,
+        atol=atol,
+    )
+    # The error for d_weight and d_bias is quite a bit higher
+    assert torch.allclose(
+        model.fc1.weight.grad,
+        model_pt_fc1.weight.grad[
+            rank * partition_out_features : (rank + 1) * partition_out_features
+        ],
+        rtol=rtol,
+        atol=atol * 10,
+    )
+    assert torch.allclose(
+        model.fc1.bias.grad,
+        model_pt_fc1.bias.grad[rank * partition_out_features : (rank + 1) * partition_out_features],
+        rtol=rtol,
+        atol=atol * 5,
+    )
+    assert torch.allclose(
+        model.fc2.weight.grad,
+        model_pt_fc2.weight.grad[
+            :, rank * partition_out_features : (rank + 1) * partition_out_features
+        ],
+        rtol=rtol,
+        atol=atol * 10,
+    )
+    if has_bias2 and rank == 0:
+        assert torch.allclose(model.fc2.bias.grad, model_pt_fc2.bias.grad, rtol=rtol, atol=atol * 5)