Add DeepSpeed (#38)

index.rst

@@ -22,6 +22,7 @@
    sources/llamafactory/index.rst
    sources/accelerate/index.rst
    sources/transformers/index.rst
+   sources/deepspeed/index.rst
    sources/onnxruntime/index.rst
    sources/open_clip/index.rst
    sources/timm/index.rst

sources/deepspeed/index.rst

@@ -0,0 +1,8 @@
+DeepSpeed
+===========
+
+.. toctree::
+   :maxdepth: 2
+
+   install.rst
+   quick_start.rst

sources/deepspeed/install.rst

@@ -0,0 +1,73 @@
+安装DeepSpeed
+==============
+
+.. note:: 
+    在本示例之前，请确保已经安装了 `昇腾环境 <../ascend/quick_install.html>`_ 和 `PyTorch <../pytorch/install.html>`_ 环境。
+
+1. 安装DeepSpeed
+-----------------
+安装DeepSpeed最简单的方式是通过 ``pip`` 。
+
+.. code-block:: shell
+    :linenos:
+
+    pip install deepspeed
+
+
+2. 通过源码安装
+------------------
+从 `GitHub <https://github.com/microsoft/DeepSpeed>`_ 克隆DeepSpeed项目后，可以通过 ``pip`` 来通过源码编译。
+
+.. code-block:: shell
+    :linenos:
+
+    pip install .
+
+
+3. 预编译DeepSpeed算子（可选）
+----------------------------------
+如果不想使用JIT编译模式，而想要预编译DeepSpeed算子，可以通过设置环境变量的方式完成算子的预编译。
+
+.. code-block:: shell
+    :linenos:
+
+    DS_BUILD_OPS=1 pip install deepspeed
+
+4. 安装验证
+-----------
+
+安装完成后，可以通过 ``ds_report`` 命令查看安装结果
+
+.. code-block:: shell
+    :linenos:
+
+    --------------------------------------------------
+    DeepSpeed C++/CUDA extension op report
+    --------------------------------------------------
+    NOTE: Ops not installed will be just-in-time (JIT) compiled at
+        runtime if needed. Op compatibility means that your system
+        meet the required dependencies to JIT install the op.
+    --------------------------------------------------
+    JIT compiled ops requires ninja
+    ninja .................. [OKAY]
+    --------------------------------------------------
+    op name ................ installed .. compatible
+    --------------------------------------------------
+    deepspeed_not_implemented  [NO] ....... [OKAY]
+    async_io ............... [NO] ....... [OKAY]
+    cpu_adagrad ............ [NO] ....... [OKAY]
+    cpu_adam ............... [NO] ....... [OKAY]
+    cpu_lion ............... [NO] ....... [OKAY]
+    fused_adam ............. [NO] ....... [OKAY]
+    transformer_inference .. [NO] ....... [OKAY]
+    --------------------------------------------------
+    DeepSpeed general environment info:
+    torch install path ............... ['/root/miniconda3/envs/ds/lib/python3.10/site-packages/torch']
+    torch version .................... 2.2.0
+    deepspeed install path ........... ['/root/miniconda3/envs/ds/lib/python3.10/site-packages/deepspeed']
+    deepspeed info ................... 0.14.4, unknown, unknown
+    deepspeed wheel compiled w. ...... torch 2.2
+    torch_npu install path ........... ['/root/miniconda3/envs/ds/lib/python3.10/site-packages/torch_npu']
+    torch_npu version ................ 2.2.0
+    ascend_cann version .............. 8.0.RC2.alpha002
+    shared memory (/dev/shm) size .... 20.00 GB

sources/deepspeed/quick_start.rst

@@ -0,0 +1,34 @@
+快速开始
+==========
+
+.. note:: 
+    在本示例之前，请确保已经安装了 `DeepSpeed <./install.html>`_ 环境。 如果还未安装，可以执行 ``pip install deepspeed`` 完成安装。
+
+
+1. 使用DeepSpeed多卡并行训练
+-------------------------------
+以下代码使用了cifar10数据集，使用DeepSpeed训练模型在多张NPU卡上进行模型训练（来自 `DeepSpeed Examples <https://github.com/microsoft/DeepSpeedExamples/blob/master/training/cifar/cifar10_deepspeed.py>`_），自DeepSpeed v0.12.6之后，代码无需任何修改，即可自动检测NPU并进行训练。
+
+.. rli:: https://raw.githubusercontent.com/microsoft/DeepSpeedExamples/master/training/cifar/cifar10_deepspeed.py
+    :language: python
+    :linenos:
+
+2. 训练结果查看
+----------------
+训练完成后，会打印模型对图像识别的结果。
+
+.. code-block:: shell
+    :linenos:
+
+    Finished Training
+    Accuracy of the network on the 10000 test images:  57 %
+    Accuracy of plane :  65 %
+    Accuracy of   car :  67 %
+    Accuracy of  bird :  52 %
+    Accuracy of   cat :  34 %
+    Accuracy of  deer :  52 %
+    Accuracy of   dog :  49 %
+    Accuracy of  frog :  59 %
+    Accuracy of horse :  66 %
+    Accuracy of  ship :  66 %
+    Accuracy of truck :  56 %