### Audio Transforms Setup

Source: https://github.com/mindspore-ai/docs/blob/master/tutorials/source_zh_cn/dataset/eager.ipynb

Initializes necessary libraries for audio transformation examples in Eager mode. Supports numpy.array as input.

```python
import numpy as np
import matplotlib.pyplot as plt
import scipy.io.wavfile as wavfile
from download import download

import mindspore.dataset as ds
import mindspore.dataset.audio as audio

ds.config.set_seed(5)
```

--------------------------------

### Ray Deployment - Master Node

Source: https://github.com/mindspore-ai/docs/blob/master/docs/vllm_mindspore/docs/source_zh_cn/user_guide/supported_features/parallel/parallel.md

Example command to start the master node using the Ray backend for distributed deployment. This simplifies the setup compared to multiprocess by managing worker processes.

```bash
# 主节点：
vllm-mindspore serve MindSpore-Lab/DeepSeek-R1-0528-A8W8 --trust-remote-code --max-num-seqs=256 --max-model-len=32768 --max-num-batched-tokens=4096 --block-size=128 --gpu-memory-utilization=0.9 --tensor-parallel-size 4 --data-parallel-size 4 --data-parallel-size-local 2 --enable-expert-parallel --additional-config '{"expert_parallel": 4}' --data-parallel-backend ray --quantization ascend
```

--------------------------------

### vLLM Service Startup Output

Source: https://github.com/mindspore-ai/docs/blob/master/docs/vllm_mindspore/docs/source_en/getting_started/tutorials/qwen2.5_32b_multiNPU/qwen2.5_32b_multiNPU.md

Example log output indicating that the vLLM service has started successfully and is ready to accept requests.

```text
INFO:   Started server process [6363]
INFO:   Waiting for application startup.
INFO:   Application startup complete.
```

--------------------------------

### Launch Multi-Node Qwen3 MOE with Expert Parallel using Ray

Source: https://github.com/mindspore-ai/docs/blob/master/docs/vllm_mindspore/docs/source_zh_cn/user_guide/supported_features/parallel/parallel.md

Example command for a two-node, four-GPU Qwen3 MOE service using Ray for multi-node expert parallel. This requires Ray environment setup and specifies the data parallel backend as Ray.

```bash
vllm-mindspore serve /path/to/Qwen3-MOE --trust-remote-code --enable-expert-parallel --additional-config '{"expert_parallel": 8} --data-parallel-backend ray'
```

--------------------------------

### Run Distributed Training Script

Source: https://github.com/mindspore-ai/docs/blob/master/tutorials/source_en/parallel/multiple_mixed.md

Execute a distributed training script using the msrun startup method. This example shows how to start an eight-card distributed training session.

```bash
bash run_sapp_mix_train.sh
```

--------------------------------

### Multiprocess Deployment - Master Node

Source: https://github.com/mindspore-ai/docs/blob/master/docs/vllm_mindspore/docs/source_zh_cn/user_guide/supported_features/parallel/parallel.md

Example command to start the master node for a multiprocess deployment. Ensure `data-parallel-address` and `data-parallel-rpc-port` are set to your environment's actual values.

```bash
# 主节点：
vllm-mindspore serve MindSpore-Lab/DeepSeek-R1-0528-A8W8 --trust-remote-code --max-num-seqs=256 --max-model-len=32768 --max-num-batched-tokens=4096 --block-size=128 --gpu-memory-utilization=0.9 --tensor-parallel-size 4 --data-parallel-size 4 --data-parallel-size-local 2 --data-parallel-start-rank 0 --data-parallel-address 127.0.0.1 --data-parallel-rpc-port 29550 --enable-expert-parallel --additional-config '{"expert_parallel": 4}' --quantization ascend
```

--------------------------------

### Launch Single-Node Qwen3 MOE with Expert Parallel

Source: https://github.com/mindspore-ai/docs/blob/master/docs/vllm_mindspore/docs/source_zh_cn/user_guide/supported_features/parallel/parallel.md

Example command to start a single-node, eight-GPU Qwen3 MOE service with expert parallel enabled. This command specifies the model path, enables trust remote code, and configures EP with 8 workers.

```bash
vllm-mindspore serve /path/to/Qwen3-MOE --trust-remote-code --enable-expert-parallel --additional-config '{"expert_parallel": 8}'
```

--------------------------------

### Install CANN Toolkit

Source: https://github.com/mindspore-ai/docs/blob/master/tutorials/source_en/orange_pi/environment_setup.md

Execute the CANN toolkit installation command. You will be prompted to confirm the installation by typing 'Y' and pressing Enter. The installation process may take 10-15 minutes.

```bash
(base) root@orangepiaipro: /home/HwHiAiUser/Downloads#./Ascend-cann-toolkit_8.3.RC1_linux-aarch64.run --install

```

--------------------------------

### Directory Structure for Startup Method Example

Source: https://github.com/mindspore-ai/docs/blob/master/tutorials/source_en/parallel/mpirun.md

This illustrates the typical directory layout for a project using mpirun for startup, including the main script, hostfile, and execution scripts.

```text
└─ sample_code
    ├─ startup_method
       ├── net.py
       ├── hostfile
       ├── run_mpirun_1.sh
       ├── run_mpirun_2.sh
    ...
```

--------------------------------

### Check scikit-learn Installation

Source: https://github.com/mindspore-ai/docs/blob/master/docs/mindquantum/docs/source_en/case_library/classification_of_iris_by_qnn.ipynb

Use this command to verify if the scikit-learn library is already installed on your system. No setup is required.

```bash
pip show scikit-learn
```

--------------------------------

### Compile and Install from Source

Source: https://github.com/mindspore-ai/docs/blob/master/docs/mindformers/docs/source_en/guide/evaluation.md

Follow these steps to clone the repository, navigate to the directory, and install from source. This method is useful for debugging and analysis.

```bash
git clone --depth 1 -b v0.4.4 https://github.com/EleutherAI/lm-evaluation-harness
cd lm-evaluation-harness
pip install -e .
```

--------------------------------

### Configure CANN Environment Variables

Source: https://github.com/mindspore-ai/docs/blob/master/docs/vllm_mindspore/docs/source_en/getting_started/installation/installation.md

Set the default installation path for CANN and source the environment setup script. This is required after CANN installation.

```bash
LOCAL_ASCEND=/usr/local/Ascend # the root directory of run package
source ${LOCAL_ASCEND}/ascend-toolkit/set_env.sh
export ASCEND_CUSTOM_PATH=${LOCAL_ASCEND}/ascend-toolkit
```

--------------------------------

### Directory Structure for Startup Method Samples

Source: https://github.com/mindspore-ai/docs/blob/master/tutorials/source_en/parallel/rank_table.md

This text-based representation shows the directory structure for sample code related to the rank table startup method. It includes Python scripts for network definition and training, shell scripts for execution, and JSON files for rank table configurations for different numbers of cards and cluster setups.

```text
└─ sample_code
    ├─ startup_method
       ├── net.py
       ├── rank_table_8pcs.json
       ├── rank_table_16pcs.json
       ├── rank_table_cross_cluster_16pcs.json
       ├── run_rank_table.sh
       ├── run_rank_table_cluster.sh
       ├── run_rank_table_cross_cluster.sh
    ...
```

--------------------------------

### Initialize Ascend Backend

Source: https://github.com/mindspore-ai/docs/blob/master/docs/lite/cloud_docs/source_en/mindir/runtime_java.md

Set up the Ascend backend for execution. This involves initializing MSContext and adding AscendDeviceInfo. Use this when leveraging Ascend NPUs for inference.

```java
MSContext context = new MSContext();
context.init();
context.addDeviceInfo(DeviceType.DT_ASCEND, false, 0);
```

--------------------------------

### Tensor Storage Offset

Source: https://github.com/mindspore-ai/docs/blob/master/tutorials/source_en/custom_program/operation/cpp_api_for_custom_ops.md

Get the offset from the start of the tensor's storage.

```APIDOC
## GET /tensor/storage_offset

### Description
Retrieves the storage offset of the tensor.

### Method
GET

### Endpoint
/tensor/storage_offset

### Parameters
None

### Request Example
None

### Response
#### Success Response (200)
- **storage_offset** (integer) - The offset from the start of storage (in terms of elements).

#### Response Example
```json
{
  "storage_offset": 0
}
```
```

--------------------------------

### Start Online Inference (Single-Card)

Source: https://github.com/mindspore-ai/docs/blob/master/docs/vllm_mindspore/docs/source_en/user_guide/supported_features/benchmark/benchmark.md

Start online inference for a single-card setup using the vllm-mindspore serve command with a specified model. Ensure the environment variable VLLM_MS_MODEL_BACKEND is set.

```bash
vllm-mindspore serve Qwen/Qwen2.5-7B-Instruct --device auto --disable-log-requests
```

--------------------------------

### Setup Loss Function and Optimizer

Source: https://github.com/mindspore-ai/docs/blob/master/docs/mindflow/docs/source_en/data_driven/2D_steady.ipynb

Prepares the loss scaler for Ascend, initializes the WaveletTransformLoss, and sets up the Adam optimizer. Learning rate scheduling is handled by get_warmup_cosine_annealing_lr.

```python
# prepare loss scaler
if use_ascend:
    from mindspore.amp import DynamicLossScaler, all_finite, init_status
    loss_scaler = DynamicLossScaler(1024, 2, 100)
else:
    loss_scaler = None
steps_per_epoch = train_dataset.get_dataset_size()
wave_loss = WaveletTransformLoss(wave_level=optimizer_params['wave_level'])
problem = SteadyFlowWithLoss(model, loss_fn=wave_loss)
# prepare optimizer
epochs = optimizer_params["epochs"]
lr = get_warmup_cosine_annealing_lr(lr_init=optimizer_params["lr"],
                                    last_epoch=epochs,
                                    steps_per_epoch=steps_per_epoch,
                                    warmup_epochs=1)
optimizer = nn.Adam(model.trainable_params() + wave_loss.trainable_params(), learning_rate=Tensor(lr))
```

--------------------------------

### Successful Installation Output Example

Source: https://github.com/mindspore-ai/docs/blob/master/docs/mindformers/docs/source_en/installation.md

This output indicates that all checks have passed and the environment is correctly set up for MindSpore Transformers.

```text
- INFO - All checks passed, used **** seconds, the environment is correctly set up!
```

--------------------------------

### Execute MindIE Startup Script

Source: https://github.com/mindspore-ai/docs/blob/master/docs/mindformers/docs/source_en/guide/deployment.md

Use this command to start MindIE with default or specified model configurations. Ensure you are in the scripts directory.

```shell
cd ./scripts
bash run_mindie.sh --model-name qwen1_5_72b --model-path /path/to/mf_model/qwen1_5_72b
```

--------------------------------

### Complete Custom Operator Usage Example

Source: https://github.com/mindspore-ai/docs/blob/master/tutorials/source_en/custom_program/operation/op_custom_prim.rst

This comprehensive example shows the full setup for a custom sine operator using the pyfunc pattern, including context setting, function definitions, operator creation, and tensor computation.

```python
import numpy as np
import mindspore as ms
from mindspore import ops

ms.set_context(mode=ms.GRAPH_MODE)
ms.set_device(device_target="CPU")


def sin_by_numpy(x):
    return np.sin(x)

def infer_shape(x):
    return x

def infer_dtype(x):
    return x

sin_by_numpy_op = ops.Custom(func=sin_by_numpy,
                                 out_shape=infer_shape,
                                 out_dtype=infer_dtype,
                                 func_type="pyfunc")
   
input_tensor = ms.Tensor([0, 1, 0.2, 0.3, 0.4], dtype=ms.float32)
result_cus = sin_by_numpy_op(input_tensor)
print(result_cus)
```

--------------------------------

### Start Scheduler Process on Node 1

Source: https://github.com/mindspore-ai/docs/blob/master/tutorials/source_en/parallel/dynamic_cluster.md

Configure environment variables to set up the scheduler and then start the training script. Ensure the scheduler IP and port are correctly set.

```bash
export MS_WORKER_NUM=8                        # Set the total number of Worker processes in the cluster to 8 (including other node processes)
export MS_SCHED_HOST=<node_1 ip address>      # Set the Scheduler IP address to the Node 1 IP address
export MS_SCHED_PORT=8118                     # Set the Scheduler port
export MS_ROLE=MS_SCHED                       # Set the startup process to the MS_SCHED role
python ./net.py > device/scheduler.log 2>&1 &    # Start training script
```

--------------------------------

### Start Online Inference (Multi-Card)

Source: https://github.com/mindspore-ai/docs/blob/master/docs/vllm_mindspore/docs/source_en/user_guide/supported_features/benchmark/benchmark.md

Start online inference for a multi-card setup using the vllm-mindspore serve command, specifying the tensor parallel size and max model length. The --trust_remote_code flag may be required.

```bash
vllm-mindspore serve Qwen/Qwen2.5-32B-Instruct --trust_remote_code --tensor-parallel-size $TENSOR_PARALLEL_SIZE --max-model-len $MAX_MODEL_LEN
```

--------------------------------

### Start Multi-Machine Training with mpirun --hostfile

Source: https://github.com/mindspore-ai/docs/blob/master/tutorials/source_en/parallel/mpirun.md

Launches distributed training using a hostfile to specify machine configurations. Requires setting DATA_PATH and passing the hostfile path as an argument.

```bash
export DATA_PATH=./MNIST_Data/train/
HOSTFILE=$1
mpirun -n 16 --hostfile $HOSTFILE --output-filename log_output --merge-stderr-to-stdout python net.py
```

--------------------------------

### Import Common MindSpore Packages

Source: https://github.com/mindspore-ai/docs/blob/master/tutorials/source_en/beginner/quick_start.md

Import necessary packages and interfaces for MindSpore examples. Ensure 'download' is installed if running as a Notebook.

```python
import mindspore
from mindspore import nn
from mindspore.dataset import vision, transforms
from mindspore.dataset import MnistDataset
```

--------------------------------

### Initial Image Transformation Setup

Source: https://github.com/mindspore-ai/docs/blob/master/docs/mindspore/source_zh_cn/faq/data_processing.md

This snippet shows an initial setup for image transformations, including adding a ToPIL operation if the input is not already a PIL Image.

```python
transforms = [
    py_vision.ToPIL(),
    py_vision.CenterCrop(375),
    py_vision.ToTensor()
]
transform = mindspore.dataset.transforms.Compose(transforms)
data1 = data1.map(operations=decode_op, input_columns=["image"])
data1 = data1.map(operations=transform, input_columns=["image"])
```

--------------------------------

### MindSpore SequentialSampler Example

Source: https://github.com/mindspore-ai/docs/blob/master/docs/mindspore/source_en/note/api_mapping/pytorch_diff/SequentialSampler.md

Illustrates sequential sampling with MindSpore's SequentialSampler, a custom dataset, and GeneratorDataset. Requires MindSpore installation.

```python
import mindspore as ms
from mindspore.dataset import SequentialSampler

class MyMapDataset():
    def __init__(self):
        super(MyMapDataset).__init__()
        self.data = [i for i in range(4)]
    def __getitem__(self, index):
        return self.data[index]
    def __len__(self):
        return len(self.data)

ds = MyMapDataset()
sampler = SequentialSampler()
dataloader = ms.dataset.GeneratorDataset(ds, column_names=["data"], sampler=sampler)

for data in dataloader:
    print(data)
# Out:
# [Tensor(shape=[], dtype=Int64, value= 0)]
# [Tensor(shape=[], dtype=Int64, value= 1)]
# [Tensor(shape=[], dtype=Int64, value= 2)]
# [Tensor(shape=[], dtype=Int64, value= 3)]
```

--------------------------------

### SSD Model Training Setup

Source: https://github.com/mindspore-ai/docs/blob/master/tutorials/source_zh_cn/cv/ssd.ipynb

Sets up the data loading, network definition, learning rate, and optimizer for SSD model training. This snippet demonstrates the initial configuration for training an SSD model.

```python
import time

from mindspore.amp import DynamicLossScaler

set_seed(1)

# load data
mindrecord_dir = "./datasets/MindRecord_COCO"
mindrecord_file = "./datasets/MindRecord_COCO/ssd.mindrecord0"

dataset = create_ssd_dataset(mindrecord_file, batch_size=5, rank=0, use_multiprocessing=True)
dataset_size = dataset.get_dataset_size()

image, get_loc, gt_label, num_matched_boxes = next(dataset.create_tuple_iterator())

# Network definition and initialization
network = SSD300Vgg16()
init_net_param(network)

# Define the learning rate
lr = Tensor(get_lr(global_step=0 * dataset_size,
                   lr_init=0.001, lr_end=0.001 * 0.05, lr_max=0.05,
                   warmup_epochs=2, total_epochs=60, steps_per_epoch=dataset_size))

# Define the optimizer
opt = nn.Momentum(filter(lambda x: x.requires_grad, network.get_parameters()), lr,
                  0.9, 0.00015, float(1024))

```

--------------------------------

### Single-Device Fine-tuning Example

Source: https://github.com/mindspore-ai/docs/blob/master/docs/mindformers/docs/source_en/feature/start_tasks.md

Execute this Python script for single-device fine-tuning. Ensure the paths to configuration and dataset are correctly specified.

```shell
python run_mindformer.py \
--register_path research/qwen2_5 \
--config research/qwen2_5/finetune_qwen2_5_0_5b_8k.yaml \
--use_parallel False \
--run_mode finetune \
--train_dataset_dir ./path/alpaca-data.mindrecord
```

--------------------------------

### Compile and Install MindSpore Quantum from Source

Source: https://github.com/mindspore-ai/docs/blob/master/docs/mindquantum/docs/source_en/mindquantum_install.md

Compile and install MindSpore Quantum locally after cloning the source code.

```bash
cd ~/mindquantum
python setup.py install --user
```

--------------------------------

### PyTorch SequentialSampler Example

Source: https://github.com/mindspore-ai/docs/blob/master/docs/mindspore/source_en/note/api_mapping/pytorch_diff/SequentialSampler.md

Demonstrates sequential sampling using PyTorch's SequentialSampler with a custom dataset and DataLoader. Requires PyTorch installation.

```python
import torch
from torch.utils.data import SequentialSampler

class MyMapDataset(torch.utils.data.Dataset):
    def __init__(self):
        super(MyMapDataset).__init__()
        self.data = [i for i in range(4)]
    def __getitem__(self, index):
        return self.data[index]
    def __len__(self):
        return len(self.data)

ds = MyMapDataset()
sampler = SequentialSampler(ds)
dataloader = torch.utils.data.DataLoader(ds, sampler=sampler)

for data in dataloader:
    print(data)
# Out:
# tensor([0])
# tensor([1])
# tensor([2])
# tensor([3])
```

--------------------------------

### TensorBoard Service Startup Message

Source: https://github.com/mindspore-ai/docs/blob/master/docs/mindformers/docs/source_en/feature/monitor.md

This message indicates that TensorBoard has started successfully. The version number, host, and port reflect your installation and configuration.

```shell
TensorBoard 2.18.0 at http://0.0.0.0:6006/ (Press CTRL+C to quit)
```

--------------------------------

### Initialize Environment Variables for MindIE and MindSpore

Source: https://github.com/mindspore-ai/docs/blob/master/docs/mindformers/docs/source_en/guide/deployment.md

Source these scripts to set up the necessary environment variables for Ascend, MindIE, and MindSpore. Ensure correct paths are used if not installed in default locations. Be mindful of potential port conflicts with MS_SCHED_PORT if other MindIE services are running.

```bash
# Ascend
source /usr/local/Ascend/ascend-toolkit/set_env.sh
# MindIE
source /usr/local/Ascend/mindie/latest/mindie-llm/set_env.sh
source /usr/local/Ascend/mindie/latest/mindie-service/set_env.sh
# MindSpore
export LCAL_IF_PORT=8129
# Networking Configuration
export MS_SCHED_HOST=127.0.0.1     # scheduler node IP address
export MS_SCHED_PORT=8090          # Scheduler node service port
```

--------------------------------

### PyTorch Inference with NPU

Source: https://github.com/mindspore-ai/docs/blob/master/docs/mindformers/docs/source_zh_cn/example/finetune_with_glm4/finetune_with_glm4.md

Example of loading a model and performing inference using PyTorch on an NPU device. Ensure PyTorch NPU adapter is installed.

```python
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch
import torch_npu  # 导入PyTorch NPU适配库

# 加载模型和分词器
model_name = "/path/to/model"
device = torch.device("npu:0")
tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(model_name, trust_remote_code=True).half().to(device)
# 将模型设置为评估模式
model.eval()
# 输入文本
input_text = "人工智能的未来发展"
# 编码输入
input_ids = tokenizer.encode(input_text, return_tensors="pt").to(model.device)
with torch.no_grad():
    output = model.generate(
        input_ids,
        max_length=100,  # 最大生成长度
        num_return_sequences=1,  # 返回的序列数
        no_repeat_ngram_size=2,  # 避免重复的n-gram
        # early_stopping=True  # 提前停止
    )

# 解码输出
generated_text = tokenizer.decode(output[0], skip_special_tokens=True)
print("生成的文本:")
print(generated_text)
```

--------------------------------

### Muon Optimizer Configuration

Source: https://github.com/mindspore-ai/docs/blob/master/docs/mindformers/docs/source_en/feature/training_hyperparameters.md

Example YAML configuration for the Muon optimizer. This setup is suitable for large-scale deep learning tasks, particularly LLM training.

```yaml
optimizer:
  type: Muon
  adamw_betas: [0.9, 0.95]
  adamw_eps: 1.e-8
  weight_decay: 0.01
  matched_adamw_rms: 0.2
  qk_clip_threshold: 100
```

--------------------------------

### PyTorch RandomRotation Example

Source: https://github.com/mindspore-ai/docs/blob/master/docs/mindspore/source_en/note/api_mapping/pytorch_diff/RandomRotation.md

Applies a random rotation to an image using PyTorch's torchvision.transforms.RandomRotation. Ensure PyTorch and torchvision are installed. The output size is printed.

```python
from download import download
from PIL import Image

url = "https://mindspore-website.obs.cn-north-4.myhuaweicloud.com/notebook/datasets/flamingos.jpg"
download(url, './flamingos.jpg', replace=True)
orig_img = Image.open('flamingos.jpg')

# PyTorch
import torch
import torchvision.transforms as T
torch.manual_seed(1)

affine_transfomer = T.RandomRotation(degrees=(30, 70), center=(0, 0))
img_torch = affine_transfomer(orig_img)
print(img_torch.size)
# Out: (471, 292)
```

--------------------------------

### Launch Training with SWAP Configuration

Source: https://github.com/mindspore-ai/docs/blob/master/docs/mindformers/docs/source_zh_cn/feature/memory_optimization.md

Execute the training script using msrun_launcher.sh, specifying the path to the custom YAML configuration file. Ensure to replace '<machine_ip>' with your local environment's IP address.

```bash
export GLOG_v=1
export MS_MEMORY_STATISTIC=1
YAML_FILE=$1 # 用户指定YAML文件路径
ROOT_PATH=`pwd`

bash ./scripts/msrun_launcher.sh "run_mindformer.py \
    --config ${ROOT_PATH}/${YAML_FILE} \
    --run_mode train \
    --use_parallel True" \
    8 8 <machine_ip> 8118 0 output/msrun False 300
```

--------------------------------

### Instantiate and Get Dataset

Source: https://github.com/mindspore-ai/docs/blob/master/tutorials/source_zh_cn/cv/fcn8s.ipynb

Define parameters for dataset creation and model training, then instantiate the SegDataset class and retrieve the processed dataset. This setup is for a segmentation task.

```python
# 定义创建数据集的参数
IMAGE_MEAN = [103.53, 116.28, 123.675]
IMAGE_STD = [57.375, 57.120, 58.395]
DATA_FILE = "dataset/dataset_fcn8s/mindname.mindrecord"

# 定义模型训练参数
train_batch_size = 4
crop_size = 512
min_scale = 0.5
max_scale = 2.0
ignore_label = 255
num_classes = 21

# 实例化Dataset
dataset = SegDataset(image_mean=IMAGE_MEAN,
                     image_std=IMAGE_STD,
                     data_file=DATA_FILE,
                     batch_size=train_batch_size,
                     crop_size=crop_size,
                     max_scale=max_scale,
                     min_scale=min_scale,
                     ignore_label=ignore_label,
                     num_classes=num_classes,
                     num_readers=2,
                     num_parallel_calls=4)

dataset = dataset.get_dataset()
```

--------------------------------

### Start MindIE with One-Click Script

Source: https://github.com/mindspore-ai/docs/blob/master/docs/mindformers/docs/source_en/guide/deployment.md

Execute the MindIE startup script from the scripts directory. Ensure you provide the model name and its path.

```shell
cd ./scripts
bash run_mindie.sh --model-name xxx --model-path /path/to/model
```

--------------------------------

### Incorrect Core Binding Configuration Example

Source: https://github.com/mindspore-ai/docs/blob/master/tutorials/source_en/parallel/msrun_launcher.md

This example is incorrect because it omits configurations for the scheduler and device0, potentially leading to misassignment of core bindings.

```bash
# wrong example
--bind_core='{"device1":["20-29", "40-49"]}'
```

--------------------------------

### Run Distributed Inference Example

Source: https://github.com/mindspore-ai/docs/blob/master/docs/lite/cloud_docs/source_en/mindir/runtime_distributed_cpp.md

Starts distributed inference by executing the compiled program for each rank. This example shows how to launch the executable for the Ascend backend, specifying the MindIR model path, rank ID, device ID, and a configuration file. It uses shell commands to run processes in the background.

```bash
# for Ascend, run the executable file for each rank using shell commands
./build/ascend_ge_distributed /your/path/to/Matmul0.mindir 0 0 ./config_file.ini &
./build/ascend_ge_distributed /your/path/to/Matmul1.mindir 1 1 ./config_file.ini
```

--------------------------------

### Setup Optimizer and Loss Function

Source: https://github.com/mindspore-ai/docs/blob/master/docs/mindflow/docs/source_en/data_driven/burgers_KNO1D.ipynb

Configures the training environment by defining the learning rate schedule, optimizer, and loss function. Includes conditional setup for mixed precision training on Ascend hardware.

```python
train_size = train_dataset.get_dataset_size()
eval_size = eval_dataset.get_dataset_size()

lr = get_warmup_cosine_annealing_lr(lr_init=optimizer_params["lr"],
                                    last_epoch=optimizer_params["epochs"],
                                    steps_per_epoch=train_size,
                                    warmup_epochs=1)
optimizer = nn.AdamWeightDecay(model.trainable_params(),
                               learning_rate=Tensor(lr),
                               weight_decay=optimizer_params["weight_decay"])
model.set_train()
loss_fn = MSELoss()
if use_ascend:
    from mindspore.amp import DynamicLossScaler, auto_mixed_precision, all_finite
    loss_scaler = DynamicLossScaler(1024, 2, 100)
    auto_mixed_precision(model, 'O3')
else:
    loss_scaler = None
```

--------------------------------

### Load Configuration and Setup Logging

Source: https://github.com/mindspore-ai/docs/blob/master/docs/mindflow/docs/source_en/data_mechanism_fusion/phympgn.ipynb

Loads configuration from a YAML file and sets up logging for training or inference. Ensure the configuration file path is correct and specify whether it's for training.

```python
from mindflow.utils import log_config, load_yaml_config, print_log
from easydict import EasyDict
import os.path as osp
from pathlib import Path


def load_config(config_file_path, train):
    config = load_yaml_config(config_file_path)
    config['train'] = train
    config = EasyDict(config)
    log_dir = './logs'
    if train:
        log_file = f'phympgn-{config.experiment_name}'
    else:
        log_file = f'phympgn-{config.experiment_name}-te'
    if not osp.exists(osp.join(log_dir, f'{log_file}.log')):
        Path(osp.join(log_dir, f'{log_file}.log')).touch()
    log_config(log_dir, log_file)
    print_log(config)
    return config
```

```python
config_file_path = 'yamls/train.yaml'
config = load_config(config_file_path=config_file_path, train=True)
```

--------------------------------

### Create a Noisy Simulator with ChannelAdder

Source: https://github.com/mindspore-ai/docs/blob/master/docs/mindquantum/docs/source_en/middle_level/noise_simulator.ipynb

Demonstrates how to initialize a Simulator with a NoiseBackend and a sequence of adders to create a noisy quantum circuit.

```python
from mindquantum.simulator import Simulator
from mindquantum.simulator.noise import NoiseBackend

noiseless_sim = Simulator('mqvector', 2)
noiseless_circ = Circuit().h(0).rx(1.0, 1).z(1, 0).measure(1)
display_svg(noiseless_circ.svg())
res1 = noiseless_sim.sampling(noiseless_circ, shots=10000)
display(res1.svg())
```

```python
noise_sim = Simulator(NoiseBackend('mqvector', 2, seq_adder))
res2 = noise_sim.sampling(noiseless_circ, shots=10000)
display(res2.svg())
display(noise_sim.backend.transform_circ(noiseless_circ).svg())
```

--------------------------------

### Run vLLM-MindSpore with Ray

Source: https://github.com/mindspore-ai/docs/blob/master/docs/vllm_mindspore/docs/source_zh_cn/getting_started/tutorials/deepseek_parallel/deepseek_r1_671b_w8a8_dp4_tp4_ep4.md

Example command to start the vLLM-MindSpore inference service using Ray for distributed execution. This command specifies model, parallelism settings, and quantization.

```bash
vllm-mindspore serve MindSpore-Lab/DeepSeek-R1-0528-A8W8 --trust-remote-code --max-num-seqs=256 --max-model-len=32768 --max-num-batched-tokens=4096 --block-size=128 --gpu-memory-utilization=0.9 --tensor-parallel-size 4 --data-parallel-size 4 --data-parallel-size-local 2 --enable-expert-parallel --additional-config '{"expert_parallel": 4}' --data-parallel-backend ray --quantization ascend
```

--------------------------------

### Configure and Create Datasets

Source: https://github.com/mindspore-ai/docs/blob/master/docs/mindflow/docs/source_zh_cn/features/solve_pinns_by_mindflow.ipynb

Sets up the configuration parameters for the geometry and sampling, then creates both the training and test datasets. The training dataset is batched for efficient processing.

```python
in_disk = {"name": "in_disk", "center_x": 0.0, "center_y": 0.0, "radius": 1.0}
out_disk = {"name": "out_disk", "center_x": 0.0, "center_y": 0.0, "radius": 2.0}
domain = {"size": 8192, "random_sampling": True, "sampler": "uniform"}
BC = {"size": 8192, "random_sampling": True, "sampler": "uniform", "with_normal": True}
data = {"domain": domain, "BC": BC}
config = {"in_disk": in_disk, "out_disk": out_disk, "data": data}

# create training dataset
dataset = create_training_dataset(config)
train_dataset = dataset.batch(batch_size=8192)

# create test dataset
inputs, label = create_test_dataset(config)

```

--------------------------------

### Instantiate Networks with Duplicate Parameters

Source: https://github.com/mindspore-ai/docs/blob/master/docs/mindspore/source_en/faq/network_compilation.md

When instantiating multiple instances of the same network (InnerNet) that contain parameters with the same name ('name_a'), an error will occur. This example shows the problematic setup.

```python
import mindspore as ms
import mindspore.nn as nn
from mindspore import Tensor, context, ParameterTuple, Parameter


context.set_context(mode=context.GRAPH_MODE)


class InnerNet(nn.Cell):
    def __init__(self):
        super(InnerNet, self).__init__()
        self.param = Parameter(Tensor([1], ms.float32), name="name_a")

    def construct(self, x):
        return x + self.param


class OutNet1(nn.Cell):
    def __init__(self, net1, net2):
        super(OutNet1, self).__init__()
        self.param1 = ParameterTuple(net1.get_parameters())
        self.param2 = ParameterTuple(net2.get_parameters())

    def construct(self, x):
        return x + self.param1[0] + self.param2[0]


net1 = InnerNet()
net2 = InnerNet()
out_net = OutNet1(net1, net2)
res = out_net(Tensor([1], ms.float32))
print("res:", res)
```

--------------------------------

### Initiating Training Tasks with msrun_launcher.sh

Source: https://github.com/mindspore-ai/docs/blob/master/docs/mindformers/docs/source_en/feature/safetensors.md

Launch distributed training tasks using the msrun_launcher.sh script. This example shows how to initiate tasks on both the master and sub-nodes.

```shell
# The first server (master node)
bash scripts/msrun_launcher.sh "run_mindformer.py \
  --config {CONFIG_PATH} \
  --run_mode train" \
  16 8 ${ip} ${port} 0 output/msrun_log False 300
# The second server (sub-node)
bash scripts/msrun_launcher.sh "run_mindformer.py \
  --config {CONFIG_PATH} \
  --run_mode train" \
  16 8 ${ip} ${port} 1 output/msrun_log False 300
```

--------------------------------

### Compile MindSpore Lite C++ Runtime

Source: https://github.com/mindspore-ai/docs/blob/master/docs/lite/cloud_docs/source_en/mindir/runtime_cpp.md

Compile the C++ runtime using CMake and Make. Ensure environment variables are set as per the quick start guide.

```bash
mkdir build && cd build
cmake ../
make
```

--------------------------------

### MindSpore RandomRotation Example

Source: https://github.com/mindspore-ai/docs/blob/master/docs/mindspore/source_en/note/api_mapping/pytorch_diff/RandomRotation.md

Applies a random rotation to an image using MindSpore's mindspore.dataset.vision.RandomRotation. Ensure MindSpore is installed and the dataset configuration is set. The output size is printed.

```python
import mindspore as ms
import mindspore.dataset.vision as vision
ms.dataset.config.set_seed(2)

affine_transfomer = vision.RandomRotation(degrees=(30, 70), center=(0, 0))
img_ms = affine_transfomer(orig_img)
print(img_ms.size)
# Out: (471, 292)
```

--------------------------------

### Directory Setup and Model Training Mode

Source: https://github.com/mindspore-ai/docs/blob/master/tutorials/source_zh_cn/generative/gan.ipynb

Sets up the necessary directories for saving checkpoints and images, and then sets both the generator and discriminator models to training mode.

```python
# 设置参数保存路径
os.makedirs(checkpoints_path, exist_ok=True)
# 设置中间过程生成图片保存路径
os.makedirs(image_path, exist_ok=True)

net_g.set_train()
net_d.set_train()
```

--------------------------------

### Multi-Machine 2-Node 8-Card Training (Node 2)

Source: https://github.com/mindspore-ai/docs/blob/master/tutorials/source_en/parallel/msrun_launcher.md

This script is executed on the second node (rank 1) of a multi-machine setup. It connects to the master address specified and starts worker processes.

```bash
EXEC_PATH=$(pwd)
if [ ! -d "${EXEC_PATH}/MNIST_Data" ]; then
    if [ ! -f "${EXEC_PATH}/MNIST_Data.zip" ]; then
        wget http://mindspore-website.obs.cn-north-4.myhuaweicloud.com/notebook/datasets/MNIST_Data.zip
    fi
    unzip MNIST_Data.zip
fi
export DATA_PATH=${EXEC_PATH}/MNIST_Data/train/

rm -rf msrun_log
mkdir msrun_log
echo "start training"

msrun --worker_num=8 --local_worker_num=4 --master_addr=<node_1 ip address> --master_port=8118 --node_rank=1 --log_dir=msrun_log --join=True --cluster_time_out=300 net.py
```

--------------------------------

### Initialize Simulator and Circuit for Bit Phase Flip

Source: https://github.com/mindspore-ai/docs/blob/master/docs/mindquantum/docs/source_en/case_library/grover_search_algorithm.ipynb

Sets up the quantum simulator and circuit for a bit phase flip operation. Requires the 'mqvector' simulator and basic quantum gates.

```python
n_qubits = 3
sim1 = Simulator('mqvector', n_qubits)

operator1 = bitphaseflip_operator([i for i in range(1, pow(2, n_qubits))], n_qubits)

circuit1 = Circuit()
circuit1 += UN(H, n_qubits)
circuit1 += operator1

sim1.apply_circuit(circuit1)

circuit1.svg()
```

--------------------------------

### Multiprocess Deployment - Worker Node

Source: https://github.com/mindspore-ai/docs/blob/master/docs/vllm_mindspore/docs/source_zh_cn/user_guide/supported_features/parallel/parallel.md

Example command to start a worker node for a multiprocess deployment. The `headless` flag is used for worker nodes. Ensure `data-parallel-address` and `data-parallel-rpc-port` match the master node.

```bash
# 从节点：
vllm-mindspore serve MindSpore-Lab/DeepSeek-R1-0528-A8W8 --trust-remote-code --max-num-seqs=256 --max-model-len=32768 --max-num-batched-tokens=4096 --block-size=128 --gpu-memory-utilization=0.9 --headless --tensor-parallel-size 4 --data-parallel-size 4 --data-parallel-size-local 2 --data-parallel-start-rank 2 --data-parallel-address 127.0.0.1 --data-parallel-rpc-port 29550 --enable-expert-parallel --additional-config '{"expert_parallel": 4}' --quantization ascend
```

--------------------------------

### Start TensorBoard Web Visualization Service

Source: https://github.com/mindspore-ai/docs/blob/master/docs/mindformers/docs/source_en/feature/monitor.md

Use this command to launch the TensorBoard service. The logdir parameter should point to the directory where TensorBoard event files are saved. Ensure the host and port are set appropriately for your network environment.

```bash
tensorboard --logdir=./worker/tensorboard/ --host=0.0.0.0 --port=6006
```

--------------------------------

### Evaluation Log Output

Source: https://github.com/mindspore-ai/docs/blob/master/docs/mindearth/docs/source_en/medium-range/graphcast.ipynb

Example log output detailing the start and end of the evaluation process, including dataset size and performance metrics like RMSE and ACC for different forecast hours.

```text
2023-08-18 08:39:45,287 - forecast.py[line:204] - INFO: ================================Start Evaluation================================
2023-08-18 08:40:41,617 - forecast.py[line:222] - INFO: test dataset size: 8
2023-08-18 08:40:41,621 - forecast.py[line:173] - INFO: t = 6 hour: 
2023-08-18 08:40:41,622 - forecast.py[line:183] - INFO:  RMSE of Z500: 99.63094225503905, T2m: 1.7807282244821678, T850: 1.1389199313716583, U10: 1.3300655484052706
2023-08-18 08:40:41,623 - forecast.py[line:184] - INFO:  ACC  of Z500: 0.9995030164718628, T2m: 0.9949086904525757, T850: 0.9965617656707764, U10: 0.9709641933441162
2023-08-18 08:40:41,624 - forecast.py[line:173] - INFO: t = 72 hour: 
2023-08-18 08:40:41,625 - forecast.py[line:183] - INFO:  RMSE of Z500: 846.2669541832905, T2m: 5.095069601461138, T850: 4.291456435667611, U10: 5.033789250954006
2023-08-18 08:40:41,627 - forecast.py[line:184] - INFO:  ACC  of Z500: 0.9656049013137817, T2m: 0.9600029587745667, T850: 0.9581822752952576, U10: 0.5923701524734497
2023-08-18 08:40:41,628 - forecast.py[line:173] - INFO: t = 120 hour: 
2023-08-18 08:40:41,629 - forecast.py[line:183] - INFO:  RMSE of Z500: 1289.3497973601945, T2m: 7.078691998772932, T850: 5.762323874978418, U10: 6.205910397891656
2023-08-18 08:40:41,629 - forecast.py[line:184] - INFO:  ACC  of Z500: 0.9226452112197876, T2m: 0.9238687753677368, T850: 0.9285233020782471, U10: 0.366882860660553
2023-08-18 08:40:41,630 - forecast.py[line:232] - INFO: ================================End Evaluation================================


```

--------------------------------

### Install MindQuantum and Chemistry Libraries

Source: https://github.com/mindspore-ai/docs/blob/master/docs/mindquantum/docs/source_en/advanced/mqchem_tutorial.ipynb

Install the necessary libraries for MindQuantum and quantum chemistry calculations.

```bash
pip install mindquantum openfermion openfermionpyscf
```

--------------------------------

### Example CheckPoint File Paths

Source: https://github.com/mindspore-ai/docs/blob/master/docs/mindformers/docs/source_en/feature/high_availability.md

Example output from the `find` command, listing the paths of generated CheckPoint files.

```text
$ find output/checkpoint/ -name '*.ckpt'
output/checkpoint/rank_2/llama2_13b_rank_2-5_1.ckpt
output/checkpoint/rank_3/llama2_13b_rank_3-5_1.ckpt
output/checkpoint/rank_0/llama2_13b_rank_0-5_1.ckpt
output/checkpoint/rank_5/llama2_13b_rank_5-5_1.ckpt
```

--------------------------------

### Multi-Node Fine-Tuning Configuration Snippet

Source: https://github.com/mindspore-ai/docs/blob/master/docs/mindformers/docs/source_en/guide/supervised_fine_tuning.md

Example of parallel configuration settings within a YAML file for multi-node, multi-NPU fine-tuning. Adjust data_parallel, model_parallel, pipeline_stage, and context_parallel as per your cluster setup.

```yaml
parallel_config:
  data_parallel: ...
  model_parallel: ...
  pipeline_stage: ...
  context_parallel: ...
```