torchtitan-npu 0.2.2.post1__tar.gz

torchtitan_npu-0.2.2.post1/LICENSE

@@ -0,0 +1,29 @@
+BSD 3-Clause License
+
+Copyright (c) 2026 Huawei Technologies Co., Ltd. All rights reserved.
+(c) Meta Platforms, Inc. and affiliates.
+
+Redistribution and use in source and binary forms, with or without modification,
+are permitted provided that the following conditions are met:
+
+1. Redistributions of source code must retain the above copyright notice,this list
+of conditions and the following disclaimer.
+
+2. 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.
+
+3. 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.

torchtitan_npu-0.2.2.post1/NOTICE

@@ -0,0 +1,128 @@
+This project constitutes a derivative work of pytorch/torchtitan.
+
+Original Project: torchtitan
+Original License: BSD-3-Clause
+Copyright (c) Meta Platforms Inc. and affiliates
+
+The original license text of torchtitan can be found below:
+```
+BSD 3-Clause License
+
+(c) Meta Platforms, Inc. and affiliates.
+
+Redistribution and use in source and binary forms, with or without modification,
+are permitted provided that the following conditions are met:
+
+1. Redistributions of source code must retain the above copyright notice,this list
+of conditions and the following disclaimer.
+
+2. 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.
+
+3. 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.
+```
+
+Original Project: pytorch
+Original License: BSD-3-Clause
+
+The original license text of pytorch can be found below:
+```
+From PyTorch:
+
+Copyright (c) 2016-     Facebook, Inc            (Adam Paszke)
+Copyright (c) 2014-     Facebook, Inc            (Soumith Chintala)
+Copyright (c) 2011-2014 Idiap Research Institute (Ronan Collobert)
+Copyright (c) 2012-2014 Deepmind Technologies    (Koray Kavukcuoglu)
+Copyright (c) 2011-2012 NEC Laboratories America (Koray Kavukcuoglu)
+Copyright (c) 2011-2013 NYU                      (Clement Farabet)
+Copyright (c) 2006-2010 NEC Laboratories America (Ronan Collobert, Leon Bottou, Iain Melvin, Jason Weston)
+Copyright (c) 2006      Idiap Research Institute (Samy Bengio)
+Copyright (c) 2001-2004 Idiap Research Institute (Ronan Collobert, Samy Bengio, Johnny Mariethoz)
+
+From Caffe2:
+
+Copyright (c) 2016-present, Facebook Inc. All rights reserved.
+
+All contributions by Facebook:
+Copyright (c) 2016 Facebook Inc.
+
+All contributions by Google:
+Copyright (c) 2015 Google Inc.
+All rights reserved.
+
+All contributions by Yangqing Jia:
+Copyright (c) 2015 Yangqing Jia
+All rights reserved.
+
+All contributions by Kakao Brain:
+Copyright 2019-2020 Kakao Brain
+
+All contributions by Cruise LLC:
+Copyright (c) 2022 Cruise LLC.
+All rights reserved.
+
+All contributions by Tri Dao:
+Copyright (c) 2024 Tri Dao.
+All rights reserved.
+
+All contributions by Arm:
+Copyright (c) 2021, 2023-2025 Arm Limited and/or its affiliates
+
+All contributions from Caffe:
+Copyright(c) 2013, 2014, 2015, the respective contributors
+All rights reserved.
+
+All other contributions:
+Copyright(c) 2015, 2016 the respective contributors
+All rights reserved.
+
+Caffe2 uses a copyright model similar to Caffe: each contributor holds
+copyright over their contributions to Caffe2. The project versioning records
+all such contribution and copyright details. If a contributor wants to further
+mark their specific copyright on a particular contribution, they should
+indicate their copyright solely in the commit message of the change when it is
+committed.
+
+All rights reserved.
+
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are met:
+
+1. Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+
+2. 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.
+
+3. Neither the names of Facebook, Deepmind Technologies, NYU, NEC Laboratories America
+   and IDIAP Research Institute 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 OWNER 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.
+```

torchtitan_npu-0.2.2.post1/PKG-INFO

@@ -0,0 +1,247 @@
+Metadata-Version: 2.4
+Name: torchtitan_npu
+Version: 0.2.2.post1
+Summary: Ascend End-to-End Large Model Training Adaptation Framework Based on torchtitan
+Author: Huawei Hisilicon
+License: BSD 3-Clause License
+        
+        Copyright (c) 2026 Huawei Technologies Co., Ltd. All rights reserved.
+        (c) Meta Platforms, Inc. and affiliates.
+        
+        Redistribution and use in source and binary forms, with or without modification,
+        are permitted provided that the following conditions are met:
+        
+        1. Redistributions of source code must retain the above copyright notice,this list
+        of conditions and the following disclaimer.
+        
+        2. 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.
+        
+        3. 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.
+        
+Project-URL: Homepage, https://gitcode.com/cann/torchtitan-npu
+Project-URL: Documentation, https://gitcode.com/cann/torchtitan-npu/tree/master/docs
+Project-URL: Issues, https://gitcode.com/cann/torchtitan-npu/issues
+Project-URL: Repository, https://gitcode.com/cann/torchtitan-npu.git
+Keywords: pytorch,training,llm,npu,ascend
+Requires-Python: >=3.10
+Description-Content-Type: text/markdown
+License-File: LICENSE
+License-File: NOTICE
+Requires-Dist: torch
+Requires-Dist: torch_npu
+Provides-Extra: testing
+Requires-Dist: pytest>=7.0; extra == "testing"
+Requires-Dist: pytest-html>=4.0; extra == "testing"
+Requires-Dist: pytest-cov>=4.0; extra == "testing"
+Requires-Dist: pytest-mock>=3.10; extra == "testing"
+Requires-Dist: hypothesis>=6.0; extra == "testing"
+Provides-Extra: dev
+Requires-Dist: pre-commit; extra == "dev"
+Requires-Dist: pytest; extra == "dev"
+Requires-Dist: pytest-cov; extra == "dev"
+Requires-Dist: expecttest; extra == "dev"
+Requires-Dist: pyrefly==0.45.1; extra == "dev"
+Dynamic: license-file
+
+<div align="center">
+
+# torchtitan-npu
+
+<h4>基于 torchtitan 的昇腾全流程大模型训练适配插件</h4>
+
+[![Documentation](https://img.shields.io/badge/docs-latest-brightgreen.svg?style=flat)](#特性支持概览)
+[![license](https://img.shields.io/badge/license-BSD_3--Clause-lightgrey.svg)](https://gitcode.com/cann/torchtitan-npu/blob/master/LICENSE)
+[![contributing](https://img.shields.io/badge/CONTRIBUTING-teal)](https://gitcode.com/cann/torchtitan-npu/blob/master/CONTRIBUTING.md)
+[![SIG](https://img.shields.io/badge/SIG-framework--adapter-yellow)](https://gitcode.com/cann/community/tree/master/CANN/sigs/framework-adapter)
+[![contributing](https://img.shields.io/badge/pypi-0.2.2-blue)](https://gitcode.com/cann/torchtitan-npu/blob/master/CONTRIBUTING.md)
+
+</div>
+
+# 简介
+
+---
+
+`torchtitan-npu`定位为`torchtitan`的昇腾（Ascend）后端扩展插件，通过即插即用的硬件亲和性优化，充分释放NPU算力，助力`PyTorch native`训练在昇腾平台无缝、高效、稳定地运行。
+
+本插件基于社区 `ModelConverter` 拓展机制构建，已支持多维度训练优化，涵盖 NPU融合算子、图优化、图下沉、**算子自动融合**、显存管理、分布式并行以及调试维测能力等等。
+
+## 社群
+[![SIG](https://img.shields.io/badge/SIG-framework--adapter-yellow)](https://gitcode.com/cann/community/tree/master/CANN/sigs/framework-adapter)
+
+SIG 例会：[sig-framework-adapter](https://meeting.osinfra.cn/cann?sig=sig-framework-adapter)
+
+# 最新消息
+
+---
+
+- [Apr. 2026]: 🚀 **【重要特性支持】算子自动融合**：基于AscendC AutoFuse的能力，支持torch.compile + Inductor后端的算子自动融合。
+- [Apr. 2026]: 🚀 **torchtitan‑npu 正式开源**：在 NPU 上支持 4D 并行等 torchtitan 原生特性，并引入 Swap Optimizer 等 NPU 亲和优化。
+
+# Roadmap
+
+---
+
+当前季度的规划见 `torchtitan-npu` [Roadmap](https://gitcode.com/cann/torchtitan-npu/issues/5)。欢迎访问。
+
+# 安装
+
+源码安装：
+
+```shell
+git clone https://gitcode.com/cann/torchtitan-npu.git
+cd torchtitan-npu
+pip install -e . ，
+```
+
+详情请参考 [部署文档](https://gitcode.com/cann/torchtitan-npu/blob/master/docs/user-guide/installation.md) 安装torchtitan-npu及其依赖。
+
+
+# 快速上手
+指导开发者快速启动大语言模型的训练任务，具体的操作请参考：
+[快速入门(基于PyTorch框架)](https://gitcode.com/cann/torchtitan-npu/blob/master/docs/user-guide/quick-start.md)
+
+
+# 特性支持概览
+
+---
+
+<table>
+  <thead>
+    <tr>
+      <th>场景</th>
+      <th>特性名称</th>
+      <th>原生支持</th>
+      <th>NPU支持</th>
+    </tr>
+  </thead>
+  <tbody>
+    <!-- 并行能力 -->
+    <tr>
+      <td rowspan="3">并行能力</td>
+      <td>4D 并行 (FSDP2/TP/CP/PP)</td>
+      <td>✅</td>
+      <td>✅</td>
+    </tr>
+    <tr>
+      <td>专家并行 (EP/ETP)</td>
+      <td>✅</td>
+      <td>✅</td>
+    </tr>
+    <tr>
+      <td><a href="https://gitcode.com/cann/torchtitan-npu/blob/master/docs/feature_guides/parallelism/custom_cp.md">自定义 CP (DeepSeek V3.2 CP/SDPA Ulysses CP)</a></td>
+      <td>❌</td>
+      <td>✅</td>
+    </tr>
+    <!-- torch.compile -->
+    <tr>
+      <td>torch.compile</td>
+      <td><a href="https://gitcode.com/cann/torchtitan-npu/blob/master/docs/feature_guides/torch_compile.md">torch.compile</a></td>
+      <td>✅</td>
+      <td>✅</td>
+    </tr>
+    <!-- 训练精度 -->
+    <tr>
+      <td rowspan="2">训练精度</td>
+      <td>MxFP8 量化</td>
+      <td>✅</td>
+      <td>✅ (Ascend 950)</td>
+    </tr>
+    <tr>
+      <td><a href="https://gitcode.com/cann/torchtitan-npu/blob/master/docs/feature_guides/low_precision_training.md">HiF8 量化</a></td>
+      <td>❌</td>
+      <td>✅ (Ascend 950)</td>
+    </tr>
+    <!-- 训练调试与监控 -->
+    <tr>
+      <td rowspan="2">训练调试与监控</td>
+      <td>分布式 Checkpoint</td>
+      <td>✅</td>
+      <td>✅</td>
+    </tr>
+    <tr>
+      <td><a href="https://gitcode.com/cann/torchtitan-npu/blob/master/docs/feature_guides/metrics_and_debugging.md">调试工具</a></td>
+      <td>✅</td>
+      <td>✅</td>
+    </tr>
+    <!-- 性能优化 -->
+    <tr>
+      <td rowspan="2">性能优化</td>
+      <td><a href="https://gitcode.com/cann/torchtitan-npu/blob/master/docs/feature_guides/swap_optimizer.md">Swap Optimizer</a></td>
+      <td>❌</td>
+      <td>✅</td>
+    </tr>
+    <tr>
+      <td><a href="https://gitcode.com/cann/torchtitan-npu/blob/master/docs/feature_guides/npu_fused_ops.md">NPU 融合算子适配</a></td>
+      <td>❌</td>
+      <td>✅</td>
+    </tr>
+  </tbody>
+</table>
+
+# 项目结构
+torchtitan-npu 充分利用了 torchtitan 提供的 ModelConverter 插件化机制。该机制介入模型定义之后、并行策略（如 TP/FSDP）应用之前，支持以非侵入式的方式，通过注册机制对特定模块进行替换或重写。基于此方案，我们实现了融合算子优化、量化支持以及优化器增强等功能。见以下项目结构：
+```
+torchtitan-npu/
+├── torchtitan_npu/     # torchtitan_npu核心源代码
+│   ├── config/         # 对Config的补丁
+│   ├── converters/      # 基于torchtitan ModelConverter机制的补丁
+│   ├── distributed/    # 自定义分布式代码
+│   ├── models/         # 基于torchtitan-npu的模型 (如Deepseek-V3.2)
+│   ├── patches/        # 其他补丁
+│   ├── tools/          # 工具补丁
+│   ├── entry.py        # 启动训练
+│   ├── train.py        # 训练主流程补丁
+│   └── __init__.py     # torchtitan-npu 插件修改注入点
+├── docs/               # 文档
+
+```
+
+# 性能基准
+
+---
+
+### 2026.04
+
+System: Atlas 800T A3
+| Model              | Number of NPUs | Precision | GBS | Local BS | Sequence Length | FSDP | TP  | PP  | CP  | EP  | Throughput (tokens/p/s) |
+| :----------------- | :------------- | :-------- | :-- | :------- | :-------------- | :--- | :-- | :-- | :-- | :-- | :----------- |
+| [Deepseek V3.2-671B](https://gitcode.com/cann/torchtitan-npu/blob/master/torchtitan_npu/models/deepseek_v32/train_configs/deepseek_v32_671b_61layers_32k_128die.toml) | 64             | BF16      | 128  | 1       | 32768           | 4    | 4   | 1   | 8  | 64  | 103           |
+| [Deepseek V3.2-671B](https://gitcode.com/cann/torchtitan-npu/blob/master/torchtitan_npu/models/deepseek_v32/train_configs/deepseek_v32_671b_61layers_4k_128die.toml) | 64             | BF16      | 512  | 1        | 4096            | 32   | 4   | 1   | 1   | 64  | 146           |
+| [Deepseek V3-671B](https://gitcode.com/cann/torchtitan-npu/blob/master/torchtitan_npu/models/deepseek_v3/train_configs/deepseek_v3_671b_61layers_4k_128die.toml)   | 64             | BF16      | 1024 | 1       | 4096            | 32   | 4   | 1   | 1   | 128  | 546          |
+| [Deepseek V3-671B + compile(Autofuse)](https://gitcode.com/cann/torchtitan-npu/blob/master/torchtitan_npu/models/deepseek_v3/train_configs/deepseek_v3_671b_61layers_4k_128die.toml)   | 64             | BF16      | 1024 | 1       | 4096            | 32   | 4   | 1   | 1   | 128  |     576      |
+ > 注：以上MoE模型的性能数据均开启负载均衡配置moe_force_load_balance=true。
+
+# 免责声明
+
+---
+
+## 致 torchtitan‑npu 使用者
+
+1. torchtitan‑npu 提供的所有内容仅供您用于非商业目的。
+2. 对于 torchtitan‑npu 测试用例以及示例文件中所涉及的各模型和数据集，平台仅用于功能测试，华为不提供任何模型权重和数据集。如您使用这些数据进行训练，请您特别注意应遵守对应模型和数据集的 License，如您因使用这些模型和数据集而产生侵权纠纷，华为不承担任何责任。
+3. 如您在使用 torchtitan‑npu 过程中，发现任何问题（包括但不限于功能问题、合规问题），请在 GitCode 提交 issue，我们将及时审视并解决。
+
+torchtitan‑npu 功能依赖的 PyTorch 等第三方开源软件，均由第三方社区提供和维护，因第三方开源软件导致的问题的修复依赖相关社区的贡献和反馈。您应理解，torchtitan‑npu 仓库不保证对第三方开源软件本身的问题进行修复，也不保证会测试、纠正所有第三方开源软件的漏洞和错误。
+
+
+# License 声明
+
+---
+
+- torchtitan‑npu 产品的使用许可证，具体请参见 [LICENSE](https://gitcode.com/cann/torchtitan-npu/blob/master/LICENSE)。
+- torchtitan‑npu 工具 docs 目录下的文档适用相应许可证，具体请参见文档目录下的 LICENSE 文件。

torchtitan_npu-0.2.2.post1/README.md

@@ -0,0 +1,187 @@
+<div align="center">
+
+# torchtitan-npu
+
+<h4>基于 torchtitan 的昇腾全流程大模型训练适配插件</h4>
+
+[![Documentation](https://img.shields.io/badge/docs-latest-brightgreen.svg?style=flat)](#特性支持概览)
+[![license](https://img.shields.io/badge/license-BSD_3--Clause-lightgrey.svg)](https://gitcode.com/cann/torchtitan-npu/blob/master/LICENSE)
+[![contributing](https://img.shields.io/badge/CONTRIBUTING-teal)](https://gitcode.com/cann/torchtitan-npu/blob/master/CONTRIBUTING.md)
+[![SIG](https://img.shields.io/badge/SIG-framework--adapter-yellow)](https://gitcode.com/cann/community/tree/master/CANN/sigs/framework-adapter)
+[![contributing](https://img.shields.io/badge/pypi-0.2.2-blue)](https://gitcode.com/cann/torchtitan-npu/blob/master/CONTRIBUTING.md)
+
+</div>
+
+# 简介
+
+---
+
+`torchtitan-npu`定位为`torchtitan`的昇腾（Ascend）后端扩展插件，通过即插即用的硬件亲和性优化，充分释放NPU算力，助力`PyTorch native`训练在昇腾平台无缝、高效、稳定地运行。
+
+本插件基于社区 `ModelConverter` 拓展机制构建，已支持多维度训练优化，涵盖 NPU融合算子、图优化、图下沉、**算子自动融合**、显存管理、分布式并行以及调试维测能力等等。
+
+## 社群
+[![SIG](https://img.shields.io/badge/SIG-framework--adapter-yellow)](https://gitcode.com/cann/community/tree/master/CANN/sigs/framework-adapter)
+
+SIG 例会：[sig-framework-adapter](https://meeting.osinfra.cn/cann?sig=sig-framework-adapter)
+
+# 最新消息
+
+---
+
+- [Apr. 2026]: 🚀 **【重要特性支持】算子自动融合**：基于AscendC AutoFuse的能力，支持torch.compile + Inductor后端的算子自动融合。
+- [Apr. 2026]: 🚀 **torchtitan‑npu 正式开源**：在 NPU 上支持 4D 并行等 torchtitan 原生特性，并引入 Swap Optimizer 等 NPU 亲和优化。
+
+# Roadmap
+
+---
+
+当前季度的规划见 `torchtitan-npu` [Roadmap](https://gitcode.com/cann/torchtitan-npu/issues/5)。欢迎访问。
+
+# 安装
+
+源码安装：
+
+```shell
+git clone https://gitcode.com/cann/torchtitan-npu.git
+cd torchtitan-npu
+pip install -e . ，
+```
+
+详情请参考 [部署文档](https://gitcode.com/cann/torchtitan-npu/blob/master/docs/user-guide/installation.md) 安装torchtitan-npu及其依赖。
+
+
+# 快速上手
+指导开发者快速启动大语言模型的训练任务，具体的操作请参考：
+[快速入门(基于PyTorch框架)](https://gitcode.com/cann/torchtitan-npu/blob/master/docs/user-guide/quick-start.md)
+
+
+# 特性支持概览
+
+---
+
+<table>
+  <thead>
+    <tr>
+      <th>场景</th>
+      <th>特性名称</th>
+      <th>原生支持</th>
+      <th>NPU支持</th>
+    </tr>
+  </thead>
+  <tbody>
+    <!-- 并行能力 -->
+    <tr>
+      <td rowspan="3">并行能力</td>
+      <td>4D 并行 (FSDP2/TP/CP/PP)</td>
+      <td>✅</td>
+      <td>✅</td>
+    </tr>
+    <tr>
+      <td>专家并行 (EP/ETP)</td>
+      <td>✅</td>
+      <td>✅</td>
+    </tr>
+    <tr>
+      <td><a href="https://gitcode.com/cann/torchtitan-npu/blob/master/docs/feature_guides/parallelism/custom_cp.md">自定义 CP (DeepSeek V3.2 CP/SDPA Ulysses CP)</a></td>
+      <td>❌</td>
+      <td>✅</td>
+    </tr>
+    <!-- torch.compile -->
+    <tr>
+      <td>torch.compile</td>
+      <td><a href="https://gitcode.com/cann/torchtitan-npu/blob/master/docs/feature_guides/torch_compile.md">torch.compile</a></td>
+      <td>✅</td>
+      <td>✅</td>
+    </tr>
+    <!-- 训练精度 -->
+    <tr>
+      <td rowspan="2">训练精度</td>
+      <td>MxFP8 量化</td>
+      <td>✅</td>
+      <td>✅ (Ascend 950)</td>
+    </tr>
+    <tr>
+      <td><a href="https://gitcode.com/cann/torchtitan-npu/blob/master/docs/feature_guides/low_precision_training.md">HiF8 量化</a></td>
+      <td>❌</td>
+      <td>✅ (Ascend 950)</td>
+    </tr>
+    <!-- 训练调试与监控 -->
+    <tr>
+      <td rowspan="2">训练调试与监控</td>
+      <td>分布式 Checkpoint</td>
+      <td>✅</td>
+      <td>✅</td>
+    </tr>
+    <tr>
+      <td><a href="https://gitcode.com/cann/torchtitan-npu/blob/master/docs/feature_guides/metrics_and_debugging.md">调试工具</a></td>
+      <td>✅</td>
+      <td>✅</td>
+    </tr>
+    <!-- 性能优化 -->
+    <tr>
+      <td rowspan="2">性能优化</td>
+      <td><a href="https://gitcode.com/cann/torchtitan-npu/blob/master/docs/feature_guides/swap_optimizer.md">Swap Optimizer</a></td>
+      <td>❌</td>
+      <td>✅</td>
+    </tr>
+    <tr>
+      <td><a href="https://gitcode.com/cann/torchtitan-npu/blob/master/docs/feature_guides/npu_fused_ops.md">NPU 融合算子适配</a></td>
+      <td>❌</td>
+      <td>✅</td>
+    </tr>
+  </tbody>
+</table>
+
+# 项目结构
+torchtitan-npu 充分利用了 torchtitan 提供的 ModelConverter 插件化机制。该机制介入模型定义之后、并行策略（如 TP/FSDP）应用之前，支持以非侵入式的方式，通过注册机制对特定模块进行替换或重写。基于此方案，我们实现了融合算子优化、量化支持以及优化器增强等功能。见以下项目结构：
+```
+torchtitan-npu/
+├── torchtitan_npu/     # torchtitan_npu核心源代码
+│   ├── config/         # 对Config的补丁
+│   ├── converters/      # 基于torchtitan ModelConverter机制的补丁
+│   ├── distributed/    # 自定义分布式代码
+│   ├── models/         # 基于torchtitan-npu的模型 (如Deepseek-V3.2)
+│   ├── patches/        # 其他补丁
+│   ├── tools/          # 工具补丁
+│   ├── entry.py        # 启动训练
+│   ├── train.py        # 训练主流程补丁
+│   └── __init__.py     # torchtitan-npu 插件修改注入点
+├── docs/               # 文档
+
+```
+
+# 性能基准
+
+---
+
+### 2026.04
+
+System: Atlas 800T A3
+| Model              | Number of NPUs | Precision | GBS | Local BS | Sequence Length | FSDP | TP  | PP  | CP  | EP  | Throughput (tokens/p/s) |
+| :----------------- | :------------- | :-------- | :-- | :------- | :-------------- | :--- | :-- | :-- | :-- | :-- | :----------- |
+| [Deepseek V3.2-671B](https://gitcode.com/cann/torchtitan-npu/blob/master/torchtitan_npu/models/deepseek_v32/train_configs/deepseek_v32_671b_61layers_32k_128die.toml) | 64             | BF16      | 128  | 1       | 32768           | 4    | 4   | 1   | 8  | 64  | 103           |
+| [Deepseek V3.2-671B](https://gitcode.com/cann/torchtitan-npu/blob/master/torchtitan_npu/models/deepseek_v32/train_configs/deepseek_v32_671b_61layers_4k_128die.toml) | 64             | BF16      | 512  | 1        | 4096            | 32   | 4   | 1   | 1   | 64  | 146           |
+| [Deepseek V3-671B](https://gitcode.com/cann/torchtitan-npu/blob/master/torchtitan_npu/models/deepseek_v3/train_configs/deepseek_v3_671b_61layers_4k_128die.toml)   | 64             | BF16      | 1024 | 1       | 4096            | 32   | 4   | 1   | 1   | 128  | 546          |
+| [Deepseek V3-671B + compile(Autofuse)](https://gitcode.com/cann/torchtitan-npu/blob/master/torchtitan_npu/models/deepseek_v3/train_configs/deepseek_v3_671b_61layers_4k_128die.toml)   | 64             | BF16      | 1024 | 1       | 4096            | 32   | 4   | 1   | 1   | 128  |     576      |
+ > 注：以上MoE模型的性能数据均开启负载均衡配置moe_force_load_balance=true。
+
+# 免责声明
+
+---
+
+## 致 torchtitan‑npu 使用者
+
+1. torchtitan‑npu 提供的所有内容仅供您用于非商业目的。
+2. 对于 torchtitan‑npu 测试用例以及示例文件中所涉及的各模型和数据集，平台仅用于功能测试，华为不提供任何模型权重和数据集。如您使用这些数据进行训练，请您特别注意应遵守对应模型和数据集的 License，如您因使用这些模型和数据集而产生侵权纠纷，华为不承担任何责任。
+3. 如您在使用 torchtitan‑npu 过程中，发现任何问题（包括但不限于功能问题、合规问题），请在 GitCode 提交 issue，我们将及时审视并解决。
+
+torchtitan‑npu 功能依赖的 PyTorch 等第三方开源软件，均由第三方社区提供和维护，因第三方开源软件导致的问题的修复依赖相关社区的贡献和反馈。您应理解，torchtitan‑npu 仓库不保证对第三方开源软件本身的问题进行修复，也不保证会测试、纠正所有第三方开源软件的漏洞和错误。
+
+
+# License 声明
+
+---
+
+- torchtitan‑npu 产品的使用许可证，具体请参见 [LICENSE](https://gitcode.com/cann/torchtitan-npu/blob/master/LICENSE)。
+- torchtitan‑npu 工具 docs 目录下的文档适用相应许可证，具体请参见文档目录下的 LICENSE 文件。

torchtitan_npu-0.2.2.post1/assets/version.txt

@@ -0,0 +1 @@
+0.2.2.post1

torchtitan_npu-0.2.2.post1/pyproject.toml

@@ -0,0 +1,80 @@
+# Copyright (c) 2026 Huawei Technologies Co., Ltd. All rights reserved.
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+# ---- All project specifications ---- #
+[project]
+name = "torchtitan_npu"
+description = "Ascend End-to-End Large Model Training Adaptation Framework Based on torchtitan"
+readme = "README.md"
+requires-python = ">=3.10"
+license = {file = "LICENSE"}
+authors = [
+    { name = "Huawei Hisilicon" },
+]
+keywords = ["pytorch", "training", "llm", "npu", "ascend"]
+dependencies = [
+    "torch",
+    "torch_npu",
+]
+dynamic = ["version"]
+
+[project.urls]
+Homepage = "https://gitcode.com/cann/torchtitan-npu"
+Documentation = "https://gitcode.com/cann/torchtitan-npu/tree/master/docs"
+Issues = "https://gitcode.com/cann/torchtitan-npu/issues"
+Repository = "https://gitcode.com/cann/torchtitan-npu.git"
+
+[project.optional-dependencies]
+testing = [
+    "pytest>=7.0",
+    "pytest-html>=4.0",
+    "pytest-cov>=4.0",
+    "pytest-mock>=3.10",
+    "hypothesis>=6.0",
+]
+dev = [
+    "pre-commit",
+    "pytest",
+    "pytest-cov",
+    "expecttest", # test_tokenizer
+    "pyrefly==0.45.1",
+]
+
+[tool.setuptools.dynamic]
+version = {file = "assets/version.txt"}
+
+[tool.codespell]
+skip = ["*.json"]
+ignore-words-list = ["cann"]
+
+# ---- Explicit project build information ---- #
+[build-system]
+requires = ["setuptools>=61.0"]
+build-backend = "setuptools.build_meta"
+
+[tool.setuptools.packages.find]
+where = [""]
+include = ["torchtitan_npu*"]
+exclude = ["CI*", "test_reports*", "cloudcache*", "third_party*"]
+
+[tool.pytest.ini_options]
+addopts = ["--showlocals", "-v", "--tb=short", "--import-mode=importlib"]
+testpaths = ["tests"]
+python_files = ["test_*.py"]
+python_classes = ["Test*"]
+python_functions = ["test_*"]
+markers = [
+    "nightly: heavy-weight validation reserved for nightly or dedicated multi-card environments",
+    "smoke: integration/e2e smoke coverage for NPU features and training paths",
+]
+
+[tool.pyrefly]
+project-excludes = ["**/tests/**", "CI"]
+ignore-missing-imports = ["torchao.*", "torchft"]  # optional dependencies
+search-path = ["../pytorch"]  # local built pytorch
+replace-imports-with-any = ["torch_npu", "torch.npu"]
+
+[tool.usort]
+known_first_party = ["torchtitan_npu", "tests"]