deploy: 14f36fb

.buildinfo

@@ -0,0 +1,4 @@
+# Sphinx build info version 1
+# This file records the configuration used when building these files. When it is not found, a full rebuild will be done.
+config: a41fc0cab2f773448e4ba86f392a4181
+tags: 645f666f9bcd5a90fca523b33c5a78b7

_sources/sources/Diffusers/index.rst.txt

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

_sources/sources/Diffusers/install.rst.txt

@@ -0,0 +1,52 @@
+安装指南
+==============
+
+本教程面向使用 Diffusers & 昇腾开发者，帮助完成昇腾环境下 Diffusers 的安装。
+
+昇腾环境安装
+------------
+
+请根据已有昇腾产品型号及CPU架构等按照 :doc:`快速安装昇腾环境指引 <../ascend/quick_install>` 进行昇腾环境安装，或直接获取对应产品的昇腾环境镜像 `ascendai/cann <https://hub.docker.com/r/ascendai/cann/tags>`_ 。
+
+.. warning::
+  CANN 最低版本为 8.0.rc1，安装 CANN 时，请同时安装 Kernel 算子包。
+
+Diffusers 安装
+------------------
+
+Python 环境创建
+------------------
+
+.. code-block:: shell
+    :linenos:
+  
+    # 创建名为 diffusers 的 python 3.10 的虚拟环境
+    conda create -y -n diffusers python=3.10
+    # 激活虚拟环境
+    conda activate diffusers
+
+
+pip 安装
+------------------
+
+通过以下指令安装 Diffusers 及 torch-npu：
+
+.. code-block:: shell
+    :linenos:
+
+    pip install diffusers torch==2.2.0 torch-npu==2.2.0 torchvision -i https://pypi.tuna.tsinghua.edu.cn/simple
+
+
+安装校验
+------------------
+
+执行以下代码，若无任何报错，仅打印模型下载过程，即说明安装成功：
+
+.. code-block:: python
+    :linenos:
+
+    from diffusers import DiffusionPipeline
+    import torch
+
+    pipeline = DiffusionPipeline.from_pretrained("runwayml/stable-diffusion-v1-5", torch_dtype=torch.float16)
+    pipeline.to("npu")

_sources/sources/Diffusers/quick_start.rst.txt

@@ -0,0 +1,102 @@
+快速开始
+==================
+
+.. note::
+    阅读本篇前，请确保已按照 :doc:`安装教程 <./install>` 准备好昇腾环境及 Diffusers ！
+
+本示例以文生图 Diffusers 库中文生图任务为样例，展示如何进行文生图模型 stable-diffusion-xl-base-1.0 的基于 LoRA 的微调及动态合并 LoRA 的推理。
+
+文生图
+-------------
+
+.. _download:
+
+模型及数据集下载
+~~~~~~~~~~~~~~~~~~~~
+
+1. 请提前下载 `stabilityai/stable-diffusion-xl-base-1.0 <https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0>`_ 模型至自定义路径
+
+2. 请提前下载 `madebyollin/sdxl-vae-fp16-fix <https://huggingface.co/madebyollin/sdxl-vae-fp16-fix>`_ 模型至自定义路径
+
+3. 请提前下载 `reach-vb/pokemon-blip-captions <https://huggingface.co/datasets/reach-vb/pokemon-blip-captions>`_ 数据集至自定义路径
+
+
+.. _finetune:
+
+基于 LoRA 的微调
+~~~~~~~~~~~~~~~~~~~~
+
+进入 Diffusers 项目目录，新建并执行以下脚本：
+
+.. note::
+
+    请根据 :ref:`download` 中模型及数据集的实际缓存路径指定 stable-diffusion-xl-base-1.0 模型缓存路径 ``MODEL_NAME``，sdxl-vae-fp16-fix 模型缓存路径 ``VAE_NAME`` 和。
+
+.. code-block:: shell
+    :linenos:
+    :emphasize-lines: 1,2,3
+
+    export MODEL_NAME="./models_ckpt/stable-diffusion-xl-base-1.0/"
+    export VAE_NAME="./ckpt/sdxl-vae-fp16-fix"
+    export TRAIN_DIR="~/diffusers/data/pokemon-blip-captions/pokemon"
+
+    python3  ./examples/text_to_image/train_text_to_image_lora_sdxl.py \
+        --pretrained_model_name_or_path=$MODEL_NAME \
+        --pretrained_vae_model_name_or_path=$VAE_NAME \
+        --dataset_name=$DATASET_NAME --caption_column="text" \
+        --resolution=1024 \
+        --random_flip \
+        --train_batch_size=1 \
+        --num_train_epochs=2 \
+        --checkpointing_steps=500 \
+        --learning_rate=1e-04 \
+        --lr_scheduler="constant" \
+        --lr_warmup_steps=0 \
+        --mixed_precision="no" \
+        --seed=42 \
+        --output_dir="sd-pokemon-model-lora-sdxl" \
+        --validation_prompt="cute dragon creature"
+
+微调过程无报错，并且终端显示 ``Steps: 100%`` 的进度条说明微调成功。
+
+
+动态合并 LoRA 的推理
+~~~~~~~~~~~~~~~~~~~~
+
+.. note::
+
+    请根据 :ref:`download` 中模型实际缓存路径指定 ``model_path``
+
+    根据  :ref:`finetune` 中指定的 LoRA 模型路径 ``output_dir`` 指定 ``lora_model_path``
+
+    [可选] 修改 ``prompt`` 可使得生成图像改变
+
+.. code-block:: python
+    :linenos:
+    :emphasize-lines: 9
+
+    from diffusers import DiffusionPipeline
+    import torch
+
+    lora_model_path = "path/to/sd-pokemon-model-lora-sdxl/checkpoint-800/"
+    model_path = "./models_ckpt/stable-diffusion-xl-base-1.0/"
+    pipe = DiffusionPipeline.from_pretrained(model_path, torch_dtype=torch.float16)
+
+    # 将模型放到 NPU 上
+    pipe.to("npu")
+
+    # 加载 LoRA 权重
+    pipe.load_lora_weights(lora_model_path)
+    # 输入 prompt
+    prompt = "Sylveon Pokemon with elegant features, magical design, \
+            light purple aura, extremely detailed and intricate markings, \
+            photo realistic, unreal engine, octane render"
+    # 推理
+    image = pipe(prompt, num_inference_steps=30, guidance_scale=7.5).images[0]
+
+    image.save("pokemon-finetuned-inference-generation.png")
+
+
+微调过程无报错，并且终端显示 ``Loading pipeline components...: 100%`` 的进度条说明微调成功。
+查看当前目录下保存的 ``pokemon-finetuned-inference-generation.png`` 图像，可根据 ``prompt`` 生成内容相关的图像说明推理成功。
+

_sources/sources/accelerate/index.rst.txt

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

_sources/sources/accelerate/install.rst.txt

@@ -0,0 +1,28 @@
+安装指南
+==============
+
+本教程面向使用 Accelerate & 昇腾的开发者，帮助完成昇腾环境下 Accelerate 的安装。
+
+Accelerate 下载安装
+--------------------
+
+.. note::
+
+    阅读本篇前，请确保已按照 :doc:`安装教程 <./install>` 准备好昇腾环境！
+    或者直接使用具备昇腾环境的镜像 `ascendai/cann:8.0.rc1-910b-ubuntu22.04 <https://hub.docker.com/layers/ascendai/cann/8.0.rc1-910b-ubuntu22.04/images/sha256-29ef8aacf6b2babd292f06f00b9190c212e7c79a947411e213135e4d41a178a9?context=explore>`_,
+    更多的版本可至 `ascendai/cann <https://hub.docker.com/r/ascendai/cann/tags>`_ 获取。
+
+启动镜像
+:::::::::::::::::
+
+.. code-block:: shell
+  
+  docker run -itd --network host -v /usr/local/dcmi:/usr/local/dcmi -v /usr/local/bin/npu-smi:/usr/local/bin/npu-smi -v /usr/local/Ascend/driver:/usr/local/Ascend/driver -v /etc/ascend_install.info:/etc/ascend_install.info --device /dev/davinci7 --device /dev/davinci_manager --device /dev/devmm_svm --device /dev/hisi_hdc --shm-size 16G --name accelerate ascendai/cann:8.0.rc1-910b-ubuntu22.04 bash
+
+安装 Accelerate 及依赖包
+::::::::::::::::::::::::::
+
+.. code-block:: shell
+
+  pip install torch==2.2.0 torch_npu==2.2.0 accelerate -i https://pypi.tuna.tsinghua.edu.cn/simple
+

_sources/sources/accelerate/quick_start.rst.txt

@@ -0,0 +1,69 @@
+快速开始
+============
+
+.. note::
+    阅读本篇前，请确保已按照 :doc:`安装指南 <./install>` 准备好昇腾环境及 Accelerate !
+
+本教程以一个简单的 NLP 模型为例，讲述如何使用 Accelerate 在昇腾 NPU 上进行模型的训练。
+
+前置准备
+------------
+
+本篇将使用到 HuggingFace 其他工具链及 scikit-learn 库，请使用以下指令安装：
+
+.. code-block::
+
+  pip install datasets evaluate transformers scikit-learn -i https://pypi.tuna.tsinghua.edu.cn/simple
+
+本篇样例代码为 Accelrate 官方样例，需提前进行下载
+
+.. code-block::
+
+  git clone https://github.com/huggingface/accelerate.git
+
+模型训练
+------------
+
+.. code-block::
+  :linenos:
+
+  # 替换HF域名，方便国内用户进行数据及模型的下载
+  export HF_ENDPOINT=https://hf-mirror.com
+  # 进入项目目录     
+  cd accelerate/examples
+  # 模型训练
+  python nlp_example.py
+
+出现如下日志代表训练成功：
+
+::
+
+    Downloading builder script: 5.75kB [00:01, 3.69kB/s]                                                                                                  
+    tokenizer_config.json: 100%|████████████████████████████████████████████████████████████████████████████████████████| 49.0/49.0 [00:00<00:00, 237kB/s]
+    config.json: 570B [00:00, 2.23MB/s]                                                                                                                   
+    vocab.txt: 79.5kB [00:12, 3.45kB/s]Error while downloading from https://hf-mirror.com/bert-base-cased/resolve/main/vocab.txt: HTTPSConnectionPool(host='hf-mirror.com', port=443): Read timed out.
+    Trying to resume download...
+    vocab.txt: 213kB [00:07, 15.5kB/s]]
+    vocab.txt: 91.4kB [00:32, 2.81kB/s]
+    tokenizer.json: 436kB [00:19, 22.8kB/s] 
+    Downloading readme: 35.3kB [00:01, 26.4kB/s]
+    Downloading data: 100%|█████████████████████████████████████████████████████████████████████████████████████████████| 649k/649k [00:02<00:00, 288kB/s]
+    Downloading data: 100%|██████████████████████████████████████████████████████████████████████████████████████████| 75.7k/75.7k [00:00<00:00, 77.8kB/s]
+    Downloading data: 100%|█████████████████████████████████████████████████████████████████████████████████████████████| 308k/308k [00:01<00:00, 204kB/s]
+    Generating train split: 100%|███████████████████████████████████████████████████████████████████████████| 3668/3668 [00:00<00:00, 27701.23 examples/s]
+    Generating validation split: 100%|████████████████████████████████████████████████████████████████████████| 408/408 [00:00<00:00, 73426.42 examples/s]
+    Generating test split: 100%|███████████████████████████████████████████████████████████████████████████| 1725/1725 [00:00<00:00, 246370.91 examples/s]
+    Map: 100%|███████████████████████████████████████████████████████████████████████████████████████████████| 3668/3668 [00:01<00:00, 3378.05 examples/s]
+    Map: 100%|█████████████████████████████████████████████████████████████████████████████████████████████████| 408/408 [00:00<00:00, 3553.72 examples/s]
+    Map: 100%|███████████████████████████████████████████████████████████████████████████████████████████████| 1725/1725 [00:00<00:00, 5109.03 examples/s]
+    model.safetensors: 100%|███████████████████████████████████████████████████████████████████████████████████████████| 436M/436M [02:42<00:00, 2.68MB/s]
+    Some weights of BertForSequenceClassification were not initialized from the model checkpoint at bert-base-cased and are newly initialized: ['classifier.bias', 'classifier.weight']
+    You should probably TRAIN this model on a down-stream task to be able to use it for predictions and inference.
+    huggingface/tokenizers: The current process just got forked, after parallelism has already been used. Disabling parallelism to avoid deadlocks...
+    To disable this warning, you can either:
+      - Avoid using `tokenizers` before the fork if possible
+      - Explicitly set the environment variable TOKENIZERS_PARALLELISM=(true | false)
+    You're using a BertTokenizerFast tokenizer. Please note that with a fast tokenizer, using the `__call__` method is faster than using a method to encode the text followed by a call to the `pad` method to get a padded encoding.
+    epoch 0: {'accuracy': 0.8014705882352942, 'f1': 0.8439306358381503}
+    epoch 1: {'accuracy': 0.8578431372549019, 'f1': 0.8975265017667845}
+    epoch 2: {'accuracy': 0.8700980392156863, 'f1': 0.9087779690189329}