### Build HCCL AllReduce Example

Source: https://github.com/ascend/pytorch/blob/master/examples/libtorch_hccl/README_HCCL.md

Navigate to the example directory, create a build folder, and use CMake to configure and build the HCCL AllReduce example. Ensure you have CMake and a C++ compiler installed.

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

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

### Full Example: Compiling and Running with MLIR Backend

Source: https://github.com/ascend/pytorch/blob/master/torch_npu/_inductor/ascend_npu_ir/README.md

This example demonstrates setting up the environment, defining a model, and compiling it using the MLIR backend via an environment variable, then capturing and printing the generated MLIR code.

```python
import os
# 环境变量调用需要在torch_npu初始化之前
os.environ['TORCHINDUCTOR_NPU_BACKEND'] = 'mlir'

import torch
import torch_npu
from torch._inductor.utils import run_and_get_code
import torch_npu._inductor
import torch
import triton

# config导入在compile执行之前
torch._inductor.config.npu_backend = "mlir"

# 定义模型
def op_calc (x , y):
    return x * y
x = torch.randn((3,), requires_grad=False, dtype=torch.float32, device="npu")
y = torch.randn((3,), requires_grad=False, dtype=torch.float32, device="npu")
std_out = op_calc(x, y)

# options调用，修改compile参数
compile_func = torch.compile(op_calc, options={"npu_backend": "mlir"})
compile_out, codes = run_and_get_code(compile_func,x,y)
print(codes[0])
```

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

### Install PyTorch for x86

Source: https://github.com/ascend/pytorch/blob/master/README.md

Use this command to install the PyTorch package for x86 systems. Ensure CANN is installed beforehand.

```shell
pip3 install torch==2.1.0+cpu  --index-url https://download.pytorch.org/whl/cpu
```

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

### Install pybind11

Source: https://github.com/ascend/pytorch/blob/master/examples/libtorch_triton_ascend/README.md

Install the pybind11 library, which is a dependency for this project.

```bash
pip install pybind11
```

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

### Run HCCL AllReduce Example

Source: https://github.com/ascend/pytorch/blob/master/examples/libtorch_hccl/README_HCCL.md

Execute the HCCL AllReduce example using the provided shell script. Make the script executable before running.

```bash
cd examples/libtorch_hccl
chmod +x libtorch_hccl.sh
./libtorch_hccl.sh
```

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

### Run HCCL AllReduce Example with Custom Process Count

Source: https://github.com/ascend/pytorch/blob/master/examples/libtorch_hccl/README_HCCL.md

Run the HCCL AllReduce example with a specified number of processes by setting the NUM_PROCESSES environment variable.

```bash
NUM_PROCESSES=4 ./libtorch_hccl.sh
```

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

### Install PyTorch for Aarch64

Source: https://github.com/ascend/pytorch/blob/master/README.md

Use this command to install the PyTorch package for Aarch64 systems. Ensure CANN is installed beforehand.

```shell
pip3 install torch==2.1.0
```

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

### Run Demo Scripts

Source: https://github.com/ascend/pytorch/blob/master/examples/libtorch_triton_ascend/README.md

Execute the provided shell script to run the C++ demo examples.

```bash
bash examples/run_demo.sh
```

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

### Install torch-npu Package

Source: https://github.com/ascend/pytorch/blob/master/README.md

Install the torch-npu package using pip. Ensure PyTorch and its dependencies are installed first.

```shell
pip3 install torch-npu==2.1.0.post17
```

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

### Initialize CANN Environment

Source: https://github.com/ascend/pytorch/blob/master/README.md

Source the CANN environment script before using Ascend NPU. Change the path if your installation differs.

```Shell
# Default path, change it if needed.
source /usr/local/Ascend/ascend-toolkit/set_env.sh
```

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

### Install torch-npu Dependencies

Source: https://github.com/ascend/pytorch/blob/master/README.md

Install necessary Python dependencies for torch-npu. These are pyyaml and setuptools.

```shell
pip3 install pyyaml
```

```shell
pip3 install setuptools
```

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

### torch.distributed.elastic.multiprocessing.start_processes

Source: https://github.com/ascend/pytorch/blob/master/docs/api/torch_npu_apis.md

Function to start multiple processes for distributed training.

```APIDOC
## torch.distributed.elastic.multiprocessing.start_processes

### Description
Function to start multiple processes for distributed training.

### Method
N/A (Function)

### Endpoint
N/A

### Parameters
N/A

### Request Example
N/A

### Response
N/A

### Error Handling
N/A
```

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

### torch.utils.cpp_extension.get_compiler_abi_compatibility_and_version

Source: https://github.com/ascend/pytorch/blob/master/docs/api/torch_npu_apis.md

Gets the compiler ABI compatibility and version for C++ extensions. It is compatible with NPU.

```APIDOC
## torch.utils.cpp_extension.get_compiler_abi_compatibility_and_version

### Description
Gets the compiler ABI compatibility and version for C++ extensions. This API is compatible with NPU.

### Compatibility
- NPU: Y

### Limitations
None specified.
```

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

### Compile torch_npu

Source: https://github.com/ascend/pytorch/blob/master/docker/builder/README.md

Compile the torch_npu library within the Docker container. This example uses Python 3.10.

```Shell
cd /home/pytorch
bash ci/build.sh --python=3.10
```

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

### torch.npu.init

Source: https://github.com/ascend/pytorch/blob/master/docs/api/torch_npu_apis.md

Initializes the NPU device.

```APIDOC
## torch.npu.init

### Description
Initializes the NPU device. This should be called before using any NPU functionalities.

### Method
NPU API call

### Endpoint
N/A (Python function)

### Parameters
None

### Request Example
```python
torch.npu.init()
```

### Response
#### Success Response
None (initializes the device)

### Response Example
```json
null
```
```

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

### CMake: Set PyTorch and NPU Library Paths

Source: https://github.com/ascend/pytorch/blob/master/examples/libtorch_npu/CMakeLists.txt

Configures the include and link directories for PyTorch and libtorch_npu. Ensure 'torch_path' and 'torch_npu_path' are set to the correct locations of your installations.

```cmake
# set the real path
set(torch_path "path/to/your/libtorch/for/pytorch")
if(NOT EXISTS ${torch_path})
  message(FATAL_ERROR "please set correct torch_path")
endif()
include_directories(${torch_path}/include)
link_directories(${torch_path}/lib)

# set the real path
set(torch_npu_path "path/to/your/libtorch_npu/for/torch_npu")
if(NOT EXISTS ${torch_npu_path})
  message(FATAL_ERROR "please set correct torch_npu_path")
endif()
include_directories(${torch_npu_path}/include)
link_directories(${torch_npu_path}/lib)
```

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

### Tensor.lerp

Source: https://github.com/ascend/pytorch/blob/master/docs/api/torch_npu_apis.md

Performs a linear interpolation between tensor `start` and `end` based on theウェイト `weight`.
`out = start + weight * (end - start)`.
Only supports fp16, fp32, int64, bool data types.

```APIDOC
## Tensor.lerp

### Description
Performs a linear interpolation between tensor `start` and `end` based on the weight `weight`.
`out = start + weight * (end - start)`.

### Supported Data Types
Only support fp16, fp32, int64, bool.
```

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

### torch.utils.tensorboard.writer.SummaryWriter.__init__

Source: https://github.com/ascend/pytorch/blob/master/docs/api/torch_npu_apis.md

The constructor for SummaryWriter is compatible with Torch NPU.

```APIDOC
## torch.utils.tensorboard.writer.SummaryWriter.__init__

### Description
Initializes the SummaryWriter.

### Compatibility
- **torch.utils.tensorboard.writer.SummaryWriter.__init__**: Y

### Limitations
None specified.
```

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

### Tensor.lerp_

Source: https://github.com/ascend/pytorch/blob/master/docs/api/torch_npu_apis.md

In-place version of Tensor.lerp. Performs a linear interpolation between tensor `start` and `end` based on theウェイト `weight`.
`input = start + weight * (end - start)`.
Only supports fp16, fp32, int64, bool data types.

```APIDOC
## Tensor.lerp_

### Description
In-place version of Tensor.lerp. Performs a linear interpolation between tensor `start` and `end` based on the weight `weight`.
`input = start + weight * (end - start)`.

### Supported Data Types
Only support fp16, fp32, int64, bool.
```

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

### Enter Docker Container

Source: https://github.com/ascend/pytorch/blob/master/ci/docker/README.md

Run a Docker container interactively, mounting the local PyTorch source code path into the container. Replace {code_path} with the actual path to your torch_npu source code.

```Shell
docker run -it -v /{code_path}/pytorch:/home/pytorch manylinux-builder:v1 bash
```

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

### Set Torch Installation Directory

Source: https://github.com/ascend/pytorch/blob/master/examples/libtorch_hccl/README_HCCL.md

Manually specify the installation directory for libtorch if automatic detection fails. This is typically the site-packages/torch directory within your Python environment.

```bash
export TORCH_INSTALL_DIR=/path/to/torch  # 通常在 site-packages/torch
export TORCH_NPU_INSTALL_DIR=/path/to/torch_npu
```

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

### Enter Docker Container

Source: https://github.com/ascend/pytorch/blob/master/docker/builder/README.md

Enter the manylinux-builder Docker container interactively. Mount the local torch_npu source code path into the container.

```Shell
docker run -it -v /{code_path}/pytorch:/home/pytorch manylinux-builder:v1 bash
# {code_path} is the torch_npu source code path
```

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

### Build TorchNPU Docker Image

Source: https://github.com/ascend/pytorch/blob/master/docker/OVERVIEW.md

Build a TorchNPU Docker image with specified versions for PyTorch, CANN, OS, and Python. Ensure all required build arguments are provided.

```bash
docker build \
  --build-arg TORCH_VERSION=2.10.0 \
  --build-arg CHIP_ARCH=a3 \
  --build-arg OS=ubuntu \
  --build-arg OS_VERSION=22.04 \
  --build-arg PY_VERSION=3.11 \
  --build-arg CANN_VERSION=9.0.0 \
  --build-arg ARCH=arm \
  --build-arg PY_TAG=cp311 \
  --build-arg TORCH_NPU_RELEASE_TAG=v26.0.0-pytorch2.10.0 \
  -t image_name:tag \
  -f Dockerfile .
```

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

### torch.distributed.checkpoint.LoadPlanner.set_up_planner

Source: https://github.com/ascend/pytorch/blob/master/docs/api/torch_npu_apis.md

This method is part of the LoadPlanner class and is used to set up the planner for loading operations.

```APIDOC
## torch.distributed.checkpoint.LoadPlanner.set_up_planner

### Description
This method is part of the LoadPlanner class and is used to set up the planner for loading operations.

### Method
Not specified in source

### Endpoint
Not specified in source

### Parameters
Not specified in source

### Request Example
Not specified in source

### Response
Not specified in source
```

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

### torch.distributed.fsdp.FullyShardedDataParallel.get_state_dict_type

Source: https://github.com/ascend/pytorch/blob/master/docs/api/torch_npu_apis.md

Gets the state dictionary type for FSDP.

```APIDOC
## torch.distributed.fsdp.FullyShardedDataParallel.get_state_dict_type

### Description
Gets the state dictionary type for FSDP.

### Method
Not specified in source

### Endpoint
Not specified in source

### Parameters
None specified in source

### Request Example
None specified in source

### Response
None specified in source
```

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

### torch.distributed.get_world_size

Source: https://github.com/ascend/pytorch/blob/master/docs/api/torch_npu_apis.md

Gets the total number of processes in the world.

```APIDOC
## torch.distributed.get_world_size

### Description
Gets the total number of processes participating in the distributed job. This is often referred to as the world size.

### Method
N/A (Function call)

### Endpoint
N/A

### Parameters
None

### Response
#### Success Response
- **world_size** (int) - The total number of processes in the distributed world.
```

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

### torch.distributed.get_rank

Source: https://github.com/ascend/pytorch/blob/master/docs/api/torch_npu_apis.md

Gets the rank of the current process in the world.

```APIDOC
## torch.distributed.get_rank

### Description
Gets the rank of the current process in the world. The rank is a unique integer identifier for each process in a distributed job.

### Method
N/A (Function call)

### Endpoint
N/A

### Parameters
None

### Response
#### Success Response
- **rank** (int) - The rank of the current process.
```

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

### torch.jit.script

Source: https://github.com/ascend/pytorch/blob/master/docs/api/torch_npu_apis.md

Details on the compatibility and limitations of torch.jit.script for NPU.

```APIDOC
## torch.jit.script

### Description
Provides information on the compatibility of `torch.jit.script` with NPU.

### Compatibility
- NPU Compatibility: Not specified

### Limitations
No specific limitations are documented for NPU usage.
```

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

### torch.distributed.checkpoint.StorageWriter.set_up_storage_writer

Source: https://github.com/ascend/pytorch/blob/master/docs/api/torch_npu_apis.md

Sets up the storage writer for checkpointing operations. This is a method within the StorageWriter class.

```APIDOC
## torch.distributed.checkpoint.StorageWriter.set_up_storage_writer

### Description
Sets up the storage writer for checkpointing operations. This is a method within the StorageWriter class.

### Method
(Not specified in source)

### Endpoint
(Not specified in source)

### Parameters
(Not specified in source)

### Request Example
(Not specified in source)

### Response
(Not specified in source)
```

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

### torch.backends.cuda.cufft_plan_cache.size

Source: https://github.com/ascend/pytorch/blob/master/docs/api/torch_npu_apis.md

Gets the size of the CUFFT plan cache.

```APIDOC
## torch.backends.cuda.cufft_plan_cache.size

### Description
Gets the current number of plans stored in the CUFFT plan cache.

### Method
`size`

### Endpoint
torch.backends.cuda.cufft_plan_cache.size

### Parameters
None

### Request Example
```python
# Example usage (conceptual)
cache_size = torch.backends.cuda.cufft_plan_cache.size
print(f"CUFFT plan cache size: {cache_size}")
```

### Response
Returns an integer representing the number of plans in the cache.
```

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

### torch.cuda.temperature

Source: https://github.com/ascend/pytorch/blob/master/docs/api/torch_npu_apis.md

Gets the current temperature of the CUDA device.

```APIDOC
## torch.cuda.temperature

### Description
Retrieves the current temperature of the CUDA device. This can be useful for monitoring hardware performance and thermal conditions.

### Method
Not specified (assumed to be a function call)

### Endpoint
Not applicable (SDK function)

### Parameters
None specified

### Request Example
```python
import torch

# Get the device temperature
temperature = torch.cuda.temperature()
print(f"Device temperature: {temperature}°C")
```

### Response
The temperature of the CUDA device, typically in degrees Celsius. The exact return type is not specified.
```

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

### torch.nn.init.kaiming_uniform_

Source: https://github.com/ascend/pytorch/blob/master/docs/api/torch_npu_apis.md

Initializes the input weights using a Kaiming uniform distribution.

```APIDOC
## torch.nn.init.kaiming_uniform_

### Description
Initializes the input weights using a Kaiming uniform distribution.

### Method
N/A (Function call)

### Endpoint
N/A

### Parameters
N/A (Function parameters not specified in source)

### Request Example
N/A

### Response
N/A
```

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

### Secondary Development with TorchNPU Image

Source: https://github.com/ascend/pytorch/blob/master/docker/OVERVIEW.md

Use an existing TorchNPU image as a base for secondary development. Add user software and dependencies using RUN commands.

```dockerfile
FROM quay.io/ascend/torch-npu:2.10.0-910b-ubuntu22.04-py3.11 

RUN apt update -y && \
    apt install gcc ...

...
```

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

### torch.jit.trace

Source: https://github.com/ascend/pytorch/blob/master/docs/api/torch_npu_apis.md

Details on the compatibility and limitations of torch.jit.trace for NPU.

```APIDOC
## torch.jit.trace

### Description
Provides information on the compatibility of `torch.jit.trace` with NPU.

### Compatibility
- NPU Compatibility: Y

### Limitations
No specific limitations are documented for NPU usage.
```

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

### torch.autograd.backward

Source: https://github.com/ascend/pytorch/blob/master/docs/api/torch_npu_apis.md

Performs a backward pass starting from the given tensors.

```APIDOC
## torch.autograd.backward

### Description
Computes the gradient of the given tensor with respect to the graph.

### Method
Not specified (assumed to be a Python function call)

### Endpoint
Not applicable (Python API)

### Parameters
Not specified in the source text.

### Request Example
```python
torch.autograd.backward(tensors, grad_tensors=None, retain_graph=None, create_graph=False)
```

### Response
Not specified in the source text.
```

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

### torch.jit.load

Source: https://github.com/ascend/pytorch/blob/master/docs/api/torch_npu_apis.md

Loads a traced or scripted module from disk.

```APIDOC
## torch.jit.load

### Description
Loads a traced or scripted module from disk.

### Method
Not applicable (Python function)

### Endpoint
Not applicable (Python function)

### Parameters
This function typically takes a file path or file-like object from which to load the module.

### Request Example
```python
# Example usage (conceptual)
# loaded_module = torch.jit.load('my_module.pt')
```

### Response
Returns the loaded traced or scripted module.
```

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

### torch.cuda.get_allocator_backend

Source: https://github.com/ascend/pytorch/blob/master/docs/api/torch_npu_apis.md

Gets the current CUDA memory allocator backend.

```APIDOC
## torch.cuda.get_allocator_backend

### Description
Gets the current CUDA memory allocator backend.

### Method
(Not specified in source)

### Endpoint
(Not specified in source)

### Parameters
(Not specified in source)

### Request Example
(Not specified in source)

### Response
(Not specified in source)
```

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

### torch.cuda.get_rng_state_all

Source: https://github.com/ascend/pytorch/blob/master/docs/api/torch_npu_apis.md

Gets the random number generator states for all devices.

```APIDOC
## torch.cuda.get_rng_state_all

### Description
Retrieves the current states of the random number generators (RNG) across all available CUDA devices. This is useful for reproducible random number generation.

### Method
Not specified (assumed to be a function call)

### Endpoint
Not applicable (SDK function)

### Parameters
None specified

### Request Example
```python
import torch

# Get RNG states for all devices
rng_states = torch.cuda.get_rng_state_all()
print(rng_states)
```

### Response
A tuple or list containing the RNG states for each CUDA device. The exact format of the states is not specified.
```

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

### torch.jit.export

Source: https://github.com/ascend/pytorch/blob/master/docs/api/torch_npu_apis.md

Details on the compatibility and limitations of torch.jit.export for NPU.

```APIDOC
## torch.jit.export

### Description
Provides information on the compatibility of `torch.jit.export` with NPU.

### Compatibility
- NPU Compatibility: Y

### Limitations
No specific limitations are documented for NPU usage.
```

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

### torch.cuda.clock_rate

Source: https://github.com/ascend/pytorch/blob/master/docs/api/torch_npu_apis.md

Gets the current clock rate of the CUDA device.

```APIDOC
## torch.cuda.clock_rate

### Description
Retrieves the current clock rate of the CUDA device. This can indicate the operational speed of the GPU cores.

### Method
Not specified (assumed to be a function call)

### Endpoint
Not applicable (SDK function)

### Parameters
None specified

### Request Example
```python
import torch

# Get the device clock rate
clock_rate = torch.cuda.clock_rate()
print(f"Device clock rate: {clock_rate} MHz")
```

### Response
The current clock rate of the CUDA device, typically in MHz. The exact return type is not specified.
```

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

### torch.optim.ASGD

Source: https://github.com/ascend/pytorch/blob/master/docs/api/torch_npu_apis.md

Documentation for the ASGD optimizer, including methods for registering hooks, saving/loading state, and performing optimization steps.

```APIDOC
## torch.optim.ASGD

### Description
Provides documentation for the ASGD optimizer.

### Methods
- **register_step_post_hook**: Registers a hook to be called after the step function.
- **register_step_pre_hook**: Registers a hook to be called before the step function.
- **state_dict**: Returns a dictionary containing the whole state of the optimizer.
- **step**: Performs a single optimization step.
- **zero_grad**: Zeros out the gradients of all parameters.
```

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

### torch.cuda.power_draw

Source: https://github.com/ascend/pytorch/blob/master/docs/api/torch_npu_apis.md

Gets the current power draw of the CUDA device.

```APIDOC
## torch.cuda.power_draw

### Description
Retrieves the current power draw of the CUDA device. This information is valuable for power consumption analysis and performance tuning.

### Method
Not specified (assumed to be a function call)

### Endpoint
Not applicable (SDK function)

### Parameters
None specified

### Request Example
```python
import torch

# Get the device power draw
power = torch.cuda.power_draw()
print(f"Device power draw: {power} Watts")
```

### Response
The current power draw of the CUDA device, typically in Watts. The exact return type is not specified.
```

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

### torch.npu.get_sync_debug_mode

Source: https://github.com/ascend/pytorch/blob/master/docs/api/torch_npu_apis.md

Gets the synchronization debug mode for NPU operations.

```APIDOC
## torch.npu.get_sync_debug_mode

### Description
Retrieves the current synchronization debug mode setting for NPU operations.

### Method
NPU API call

### Endpoint
N/A (Python function)

### Parameters
None

### Request Example
```python
mode = torch.npu.get_sync_debug_mode()
print(f"Sync debug mode: {mode}")
```

### Response
#### Success Response
- **mode** (int) - The current synchronization debug mode.

### Response Example
```json
0
```
```

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

### torch.distributed.rpc.RpcBackendOptions.init_method

Source: https://github.com/ascend/pytorch/blob/master/docs/api/torch_npu_apis.md

Sets or retrieves the initialization method for the RPC backend.

```APIDOC
## torch.distributed.rpc.RpcBackendOptions.init_method

### Description
Configures the method used to initialize the RPC backend.

### Method
Attribute Access (Setter/Getter)

### Endpoint
N/A (Python API)

### Parameters
- **init_method** (string): The initialization method string (e.g., 'env://').

### Response
- **init_method** (string): The currently set initialization method.
```

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

### torch.get_num_threads

Source: https://github.com/ascend/pytorch/blob/master/docs/api/torch_npu_apis.md

Gets the number of threads used for parallel operations.

```APIDOC
## torch.get_num_threads

### Description
Gets the number of threads used for parallel operations.

### Method
Not specified

### Endpoint
Not specified

### Parameters
Not specified

### Request Example
Not specified

### Response
Not specified
```

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

### torch.distributions.bernoulli.Bernoulli.sample

Source: https://github.com/ascend/pytorch/blob/master/docs/api/torch_npu_apis.md

Samples from the Bernoulli distribution.

```APIDOC
## torch.distributions.bernoulli.Bernoulli.sample

### Description
Samples from the Bernoulli distribution.

### Method
Not applicable (method call)

### Endpoint
Not applicable

### Parameters
Not applicable

### Request Example
Not applicable

### Response
#### Success Response
- **sample** (Tensor) - A sample from the distribution.
```

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

### torch.get_num_interop_threads

Source: https://github.com/ascend/pytorch/blob/master/docs/api/torch_npu_apis.md

Gets the number of interop threads used for parallel operations.

```APIDOC
## torch.get_num_interop_threads

### Description
Gets the number of interop threads used for parallel operations.

### Method
Not specified

### Endpoint
Not specified

### Parameters
Not specified

### Request Example
Not specified

### Response
Not specified
```

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

### torch.fx.GraphModule.__init__

Source: https://github.com/ascend/pytorch/blob/master/docs/api/torch_npu_apis.md

The constructor for torch.fx.GraphModule is available and compatible.

```APIDOC
## torch.fx.GraphModule.__init__

### Description
Initializes a GraphModule.

### Compatibility
Compatible (Y)
```

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

### torch.distributed.checkpoint.StorageReader.set_up_storage_reader

Source: https://github.com/ascend/pytorch/blob/master/docs/api/torch_npu_apis.md

Sets up the storage reader. This method is part of the StorageReader class in PyTorch distributed checkpointing.

```APIDOC
## torch.distributed.checkpoint.StorageReader.set_up_storage_reader

### Description
Sets up the storage reader. This method is part of the StorageReader class in PyTorch distributed checkpointing.

### Method
Not specified (assumed to be a Python method call)

### Endpoint
Not applicable (Python API)

### Parameters
Not specified in the source text.

### Request Example
Not specified in the source text.

### Response
Not specified in the source text.
```

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

### torch.profiler.schedule

Source: https://github.com/ascend/pytorch/blob/master/docs/api/torch_npu_apis.md

Defines a schedule for profiling, specifying when to start and stop profiling activities.

```APIDOC
## torch.profiler.schedule

### Description
Defines a schedule for profiling, specifying when to start and stop profiling activities.

### Method
N/A (Function/Class)

### Endpoint
N/A

### Parameters
N/A

### Request Example
N/A

### Response
N/A

### Error Handling
N/A
```

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

### torch.compiler.list_backends

Source: https://github.com/ascend/pytorch/blob/master/docs/api/torch_npu_apis.md

Lists the available compiler backends for NPU.

```APIDOC
## torch.compiler.list_backends

### Description
Lists the available compiler backends for NPU.

### Method
Not specified (assumed to be a function call)

### Endpoint
Not applicable (SDK function)

### Parameters
None explicitly documented.

### Request Example
```python
torch.compiler.list_backends()
```

### Response
None explicitly documented.
```

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

### torch.profiler.ProfilerAction

Source: https://github.com/ascend/pytorch/blob/master/docs/api/torch_npu_apis.md

Represents an action to be taken during profiling, such as starting or stopping data collection.

```APIDOC
## torch.profiler.ProfilerAction

### Description
Represents an action to be taken during profiling, such as starting or stopping data collection.

### Method
N/A (Enum/Class)

### Endpoint
N/A

### Parameters
N/A

### Request Example
N/A

### Response
N/A

### Error Handling
N/A
```

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

### torch.distributed.elastic.metrics.prof

Source: https://github.com/ascend/pytorch/blob/master/docs/api/torch_npu_apis.md

Starts or stops profiling for metrics within the distributed elastic module.

```APIDOC
## torch.distributed.elastic.metrics.prof

### Description
Starts or stops profiling for metrics within the distributed elastic module.

### Method
Not specified (likely a Python function call)

### Endpoint
Not applicable (Python API)

### Parameters
Not specified in the source text.

### Request Example
Not applicable.

### Response
Not specified in the source text.
```

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

### torch.nn.ReplicationPad3d

Source: https://github.com/ascend/pytorch/blob/master/docs/api/torch_npu_apis.md

Documentation for the ReplicationPad3d module in torch.nn.

```APIDOC
## torch.nn.ReplicationPad3d

### Description
Documentation for the ReplicationPad3d module in torch.nn.

### Method
N/A

### Endpoint
N/A

### Parameters
N/A

### Request Example
N/A

### Response
N/A
```

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

### CMakeLists.txt for LibTorch ResNet with NPU

Source: https://github.com/ascend/pytorch/blob/master/examples/libtorch_resnet/CMakeLists.txt

This CMakeLists.txt file sets up the build environment for a LibTorch ResNet project. It finds the required Torch and Torch NPU packages, configures C++ flags, defines the executable, and links the necessary libraries. Ensure the libtorch_npu directory is correctly placed relative to the CMakeLists.txt file.

```cmake
cmake_minimum_required(VERSION 3.18 FATAL_ERROR)
project(libtorch_resnet)

set(CMAKE_PREFIX_PATH ${CMAKE_PREFIX_PATH} ${CMAKE_CURRENT_SOURCE_DIR}/../../libtorch_npu/)

find_package(Torch REQUIRED)
find_package(Torch_npu REQUIRED)
set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} ${TORCH_CXX_FLAGS}")

add_executable(libtorch_resnet libtorch_resnet.cpp)
target_link_libraries(libtorch_resnet "${TORCH_LIBRARIES}")
target_link_libraries(libtorch_resnet "${TORCH_NPU_LIBRARIES}")

set_property(TARGET libtorch_resnet PROPERTY CXX_STANDARD 17)
```

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

### torch.fx.Interpreter.get_attr

Source: https://github.com/ascend/pytorch/blob/master/docs/api/torch_npu_apis.md

Gets an attribute within the interpreter. This method is part of the torch.fx.Interpreter class.

```APIDOC
## torch.fx.Interpreter.get_attr

### Description
Gets an attribute within the interpreter. This method is part of the torch.fx.Interpreter class.

### Method
N/A (Method within a Python class)

### Endpoint
N/A (Python method)

### Parameters
N/A (Details not provided in source)

### Request Example
N/A

### Response
N/A

```

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

### torch.distributed.get_group_rank

Source: https://github.com/ascend/pytorch/blob/master/docs/api/torch_npu_apis.md

Gets the rank of the current process within a specified process group.

```APIDOC
## torch.distributed.get_group_rank

### Description
Gets the rank of the current process within a specified process group.

### Method
Not specified (function call)

### Endpoint
Not applicable (function call)

### Parameters
Not specified in source.

### Response
Not specified in source.
```

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

### torch.nn.ReflectionPad3d

Source: https://github.com/ascend/pytorch/blob/master/docs/api/torch_npu_apis.md

ReflectionPad3d is supported on NPU.

```APIDOC
## torch.nn.ReflectionPad3d

### Description
Pads the input tensor using reflection of the input along the dimension.
```

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

### Node.name

Source: https://github.com/ascend/pytorch/blob/master/docs/api/torch_npu_apis.md

Gets the name of a Node in the computation graph. This is a property of a Node object.

```APIDOC
## Node.name

### Description
Retrieves the identifier or name associated with a specific node within the computational graph.

### Method
Not applicable (Python attribute/property)

### Endpoint
Not applicable (Python attribute/property)

### Parameters
None.

### Request Example
```python
# Example usage (conceptual)
# node_name = node.name
```

### Response
string: The name of the node.
```

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

### torch.distributed.checkpoint.SavePlanner.set_up_planner

Source: https://github.com/ascend/pytorch/blob/master/docs/api/torch_npu_apis.md

Sets up the planner for saving operations within the distributed checkpointing module. This method is part of the SavePlanner class.

```APIDOC
## torch.distributed.checkpoint.SavePlanner.set_up_planner

### Description
Sets up the planner for saving operations within the distributed checkpointing module. This method is part of the SavePlanner class.

### Method
Not specified in source.

### Endpoint
Not specified in source.

### Parameters
Not specified in source.

### Request Example
Not specified in source.

### Response
Not specified in source.

ERROR HANDLING:
Not specified in source.
```

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

### torch.ao.quantization.fx.custom_config.PrepareCustomConfig.from_dict

Source: https://github.com/ascend/pytorch/blob/master/docs/api/torch_npu_apis.md

Creates a PrepareCustomConfig object from a dictionary.

```APIDOC
## torch.ao.quantization.fx.custom_config.PrepareCustomConfig.from_dict

### Description
Creates a PrepareCustomConfig object from a dictionary.

### Method
Not specified (likely a static or class method)

### Endpoint
Not applicable (Python API)

### Parameters
Not specified in the source text.

### Request Example
Not applicable.

### Response
Not specified in the source text.

ERROR HANDLING:
Not specified in the source text.
```

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

### torch.nn.ZeroPad1d

Source: https://github.com/ascend/pytorch/blob/master/docs/api/torch_npu_apis.md

Documentation for the ZeroPad1d module in torch.nn.

```APIDOC
## torch.nn.ZeroPad1d

### Description
Documentation for the ZeroPad1d module in torch.nn.

### Method
N/A

### Endpoint
N/A

### Parameters
N/A

### Request Example
N/A

### Response
N/A
```

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

### Find Torch Package

Source: https://github.com/ascend/pytorch/blob/master/CMakeLists.txt

Locates the required Torch package using a specified installation directory.

```cmake
set(Torch_DIR ${PYTORCH_INSTALL_DIR}/share/cmake/Torch)
FIND_PACKAGE(Torch REQUIRED)
```

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

### torch.jit.ScriptModule Methods

Source: https://github.com/ascend/pytorch/blob/master/docs/api/torch_npu_apis.md

This section details various methods available for torch.jit.ScriptModule, which are essential for working with JIT-compiled models on NPU.

```APIDOC
## torch.jit.ScriptModule

### Description
Represents a PyTorch module that has been compiled using the TorchScript JIT compiler.

### Methods

- **add_module**: Adds a named submodule to the module.
- **apply**: Applies a function recursively to all submodules and parameters.
- **bfloat16**: Moves all parameters and buffers to bfloat16.
- **buffers**: Returns an iterator over module buffers.
- **children**: Returns an iterator over immediate children modules.
- **code**: Returns the Python source code of the module.
- **code_with_constants**: Returns the Python source code and constants of the module.
- **compile**: Compiles the module into TorchScript.
- **cpu**: Moves all parameters and buffers to the CPU.
- **cuda**: Moves all parameters and buffers to the CUDA device.
```

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

### torch.jit.fork

Source: https://github.com/ascend/pytorch/blob/master/docs/api/torch_npu_apis.md

Details on the compatibility and limitations of torch.jit.fork for NPU.

```APIDOC
## torch.jit.fork

### Description
Provides information on the compatibility of `torch.jit.fork` with NPU.

### Compatibility
- NPU Compatibility: Not specified

### Limitations
No specific limitations are documented for NPU usage.
```

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

### torch.npu.get_rng_state

Source: https://github.com/ascend/pytorch/blob/master/docs/api/torch_npu_apis.md

Gets the current random number generator state for NPU. This is the NPU equivalent of torch.cuda.get_rng_state.

```APIDOC
## torch.npu.get_rng_state

### Description
Gets the current random number generator state for NPU. This is the NPU equivalent of torch.cuda.get_rng_state.

### Method
Not Applicable (Python function)

### Endpoint
Not Applicable (Python function)

### Parameters
This function does not explicitly list parameters in the provided documentation.

### Request Example
```python
# Example usage
# rng_state = torch.npu.get_rng_state()
```

### Response
Returns the current state of the NPU random number generator.
```

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

### Build Triton Runtime Library

Source: https://github.com/ascend/pytorch/blob/master/examples/libtorch_triton_ascend/README.md

Compile the Triton runtime shared library. Navigate to the triton_runtime directory and execute the build script.

```bash
cd triton_runtime
bash build.sh
cd ..
```

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

### torch.distributed.get_global_rank

Source: https://github.com/ascend/pytorch/blob/master/docs/api/torch_npu_apis.md

Gets the global rank of the current process. This is the rank of the process within the entire distributed job.

```APIDOC
## torch.distributed.get_global_rank

### Description
Gets the global rank of the current process. This is the rank of the process within the entire distributed job.

### Method
Not specified (function call)

### Endpoint
Not applicable (function call)

### Parameters
Not specified in source.

### Response
Not specified in source.
```

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

### torch.__config__.show

Source: https://github.com/ascend/pytorch/blob/master/docs/api/torch_npu_apis.md

Displays the current configuration settings for PyTorch, indicating compatibility with NPU.

```APIDOC
## torch.__config__.show

### Description
Displays the current configuration settings for PyTorch.

### Method
```python
torch.__config__.show()
```

### Parameters
None

### Response
Outputs configuration details.

### Example
```python
torch.__config__.show()
```
```

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

### torch.nn.ZeroPad3d

Source: https://github.com/ascend/pytorch/blob/master/docs/api/torch_npu_apis.md

Documentation for the ZeroPad3d module in torch.nn.

```APIDOC
## torch.nn.ZeroPad3d

### Description
Documentation for the ZeroPad3d module in torch.nn.

### Method
N/A

### Endpoint
N/A

### Parameters
N/A

### Request Example
N/A

### Response
N/A
```

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

### torch.distributed.fsdp.FullyShardedDataParallel.state_dict_type

Source: https://github.com/ascend/pytorch/blob/master/docs/api/torch_npu_apis.md

Gets the current state dictionary type for FullyShardedDataParallel. This indicates how the state is managed (e.g., sharded, replicated).

```APIDOC
## torch.distributed.fsdp.FullyShardedDataParallel.state_dict_type

### Description
Gets the current state dictionary type for FullyShardedDataParallel. This property reflects the configuration of how the model's state is stored and accessed, which is essential for understanding checkpointing behavior.

### Method
Not specified (assumed to be a method call within a Python script)

### Endpoint
N/A (Python API)

### Parameters
N/A (Python API)

### Request Example
N/A (Python API)

### Response
N/A (Python API)
```

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

### EtcdStore

Source: https://github.com/ascend/pytorch/blob/master/docs/api/torch_npu_apis.md

An Etcd-based store for distributed coordination. Provides methods for adding, checking, getting, setting, and waiting for data.

```APIDOC
## torch.distributed.elastic.rendezvous.etcd_store.EtcdStore

### Description
An implementation of a distributed store using Etcd, used for rendezvous coordination.

### Methods
- **add**: Adds data to the store.
- **check**: Checks for the existence of data in the store.
- **get**: Retrieves data from the store.
- **set**: Sets data in the store.
- **wait**: Waits for specific conditions or data in the store.

(Detailed method signatures and parameters are not provided in the source.)
```

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

### torch.package.PackageExporter.__init__

Source: https://github.com/ascend/pytorch/blob/master/docs/api/torch_npu_apis.md

Initializes a PackageExporter object.

```APIDOC
## torch.package.PackageExporter.__init__

### Description
Constructor for the PackageExporter class. Initializes the exporter with necessary configurations.

### Compatibility
Y

### Limitations
None specified.
```

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

### torch.jit.ScriptModule Methods

Source: https://github.com/ascend/pytorch/blob/master/docs/api/torch_npu_apis.md

This section details various methods available for torch.jit.ScriptModule objects, indicating their NPU compatibility.

```APIDOC
## torch.jit.ScriptModule.inlined_graph

### Description
Provides access to the inlined graph of a ScriptModule.

### Method
N/A (Attribute/Property)

### Endpoint
N/A

### Parameters
N/A

### Request Example
N/A

### Response
N/A

## torch.jit.ScriptModule.ipu

### Description
Indicates if the ScriptModule is compatible with or utilizes IPU (Intelligence Processing Unit) features.

### Method
N/A (Attribute/Property)

### Endpoint
N/A

### Parameters
N/A

### Request Example
N/A

### Response
N/A

## torch.jit.ScriptModule.load_state_dict

### Description
Loads the state dictionary into the ScriptModule.

### Method
N/A (Method)

### Endpoint
N/A

### Parameters
N/A

### Request Example
N/A

### Response
N/A

## torch.jit.ScriptModule.modules

### Description
Returns an iterator over all submodules in the module.

### Method
N/A (Method)

### Endpoint
N/A

### Parameters
N/A

### Request Example
N/A

### Response
N/A

## torch.jit.ScriptModule.named_buffers

### Description
Returns an iterator over module buffers, yielding both the name of the buffer as well as the buffer itself.

### Method
N/A (Method)

### Endpoint
N/A

### Parameters
N/A

### Request Example
N/A

### Response
N/A

## torch.jit.ScriptModule.named_children

### Description
Returns an iterator over immediate children of this module, yielding both the name of the child as well as the child module.

### Method
N/A (Method)

### Endpoint
N/A

### Parameters
N/A

### Request Example
N/A

### Response
N/A

## torch.jit.ScriptModule.named_modules

### Description
Returns an iterator over all modules in the module, yielding both the name of the module, as well as the module itself.

### Method
N/A (Method)

### Endpoint
N/A

### Parameters
N/A

### Request Example
N/A

### Response
N/A

## torch.jit.ScriptModule.named_parameters

### Description
Returns an iterator over module parameters, yielding both the name of the parameter as well as the parameter itself.

### Method
N/A (Method)

### Endpoint
N/A

### Parameters
N/A

### Request Example
N/A

### Response
N/A

## torch.jit.ScriptModule.parameters

### Description
Returns an iterator over module parameters.

### Method
N/A (Method)

### Endpoint
N/A

### Parameters
N/A

### Request Example
N/A

### Response
N/A

## torch.jit.ScriptModule.register_backward_hook

### Description
Registers a backward hook on the module.

### Method
N/A (Method)

### Endpoint
N/A

### Parameters
N/A

### Request Example
N/A

### Response
N/A
```

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

### Tensor.int_repr

Source: https://github.com/ascend/pytorch/blob/master/docs/api/torch_npu_apis.md

Returns the integer representation of a tensor. This method provides a way to get the raw integer value of tensor elements.

```APIDOC
## Tensor.int_repr

### Description
Returns the integer representation of a tensor. This method provides a way to get the raw integer value of tensor elements.

### Method
Not specified, typically a method call on a Tensor object.

### Endpoint
Not applicable (SDK method).

### Parameters
Not specified in the source text.

### Request Example
Not specified in the source text.

### Response
Not specified in the source text.
```

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

### torch.fx.wrap

Source: https://github.com/ascend/pytorch/blob/master/docs/api/torch_npu_apis.md

torch.fx.wrap is available and compatible.

```APIDOC
## torch.fx.wrap

### Description
Utility for wrapping objects in PyTorch FX.

### Compatibility
Compatible (Y)
```

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

### torch.nn.init.kaiming_normal_

Source: https://github.com/ascend/pytorch/blob/master/docs/api/torch_npu_apis.md

Initializes the input weights using a Kaiming normal distribution.

```APIDOC
## torch.nn.init.kaiming_normal_

### Description
Initializes the input weights using a Kaiming normal distribution.

### Method
N/A (Function call)

### Endpoint
N/A

### Parameters
N/A (Function parameters not specified in source)

### Request Example
N/A

### Response
N/A
```

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

### torch.nn.SyncBatchNorm.convert_sync_batchnorm

Source: https://github.com/ascend/pytorch/blob/master/docs/api/torch_npu_apis.md

A utility function to convert existing BatchNorm layers to SyncBatchNorm layers. This is useful for migrating models to a distributed training setup.

```APIDOC
## torch.nn.SyncBatchNorm.convert_sync_batchnorm

### Description
Converts existing BatchNorm layers to SyncBatchNorm layers, facilitating migration to distributed training environments.

### Method
N/A (Static method or class method)

### Endpoint
N/A

### Parameters
N/A

### Request Example
N/A

### Response
N/A
```

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

### Compile torch_npu for Python 3.9

Source: https://github.com/ascend/pytorch/blob/master/ci/docker/README.md

Compile torch_npu within the Docker container. This example uses Python 3.9. Ensure you are in the /home/pytorch directory.

```Shell
cd /home/pytorch
bash ci/build.sh --python=3.9
```

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

### torch.ao.quantization.prepare_qat

Source: https://github.com/ascend/pytorch/blob/master/docs/api/torch_npu_apis.md

Prepares a PyTorch model for Quantization-Aware Training (QAT).

```APIDOC
## torch.ao.quantization.prepare_qat

### Description
Prepares a PyTorch model for Quantization-Aware Training (QAT).

### Method
Not applicable (Python function)

### Endpoint
Not applicable (Python function)

### Parameters
This function prepares the model for QAT. Specific parameters are not detailed in the source.

### Request Example
Not applicable (Python function)

### Response
Returns the QAT-prepared PyTorch model.

#### Success Response (200)
Not applicable (Python function)

#### Response Example
Not applicable (Python function)
```

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

### Check for libpython*.so

Source: https://github.com/ascend/pytorch/blob/master/examples/libtorch_triton_ascend/README.md

Verify the presence of the shared Python library, which is required for compilation. This command lists the library in the Python installation's lib directory.

```bash
ls $(python3 -c "import sysconfig; print(sysconfig.get_config_var('LIBDIR'))")/libpython*.so
```

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

### torch.nn.ConstantPad3d

Source: https://github.com/ascend/pytorch/blob/master/docs/api/torch_npu_apis.md

Documentation for the ConstantPad3d module in torch.nn.

```APIDOC
## torch.nn.ConstantPad3d

### Description
Documentation for the ConstantPad3d module in torch.nn.

### Method
N/A

### Endpoint
N/A

### Parameters
N/A

### Request Example
N/A

### Response
N/A
```

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

### torch.jit.trace_module

Source: https://github.com/ascend/pytorch/blob/master/docs/api/torch_npu_apis.md

Details on the compatibility and limitations of torch.jit.trace_module for NPU.

```APIDOC
## torch.jit.trace_module

### Description
Provides information on the compatibility of `torch.jit.trace_module` with NPU.

### Compatibility
- NPU Compatibility: Y

### Limitations
No specific limitations are documented for NPU usage.
```

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

### PyTorch Ascend NPU Project Architecture

Source: https://github.com/ascend/pytorch/blob/master/CONTRIBUTING.md

Provides an overview of the directory structure for the PyTorch Ascend NPU project, illustrating the organization of core modules, tools, and examples.

```text
pytorch
├── docs/                           # 项目文档
├── ci/                             # CI 构建脚本
├── tools/                          # 开发工具
├── cmake/                          # CMake 配置
├── torch_npu/                      # NPU 核心适配模块
│   ├── csrc/                       # C++ 后端实现
│   ├── distributed/                # 分布式 Python 接口
│   ├── _inductor/                  # Inductor 后端适配
│   ├── dynamo/                     # Dynamo 编译器适配
│   ├── npu/                        # NPU Python 接口
│   ├── profiler/                   # 性能分析 Python 接口
│   ├── _afd/                       # AFD Python 接口
│   ├── _logging/                   # 日志模块 Python 接口
│   ├── asd/                        # 异步检测工具
│   ├── contrib/                    # 贡献的扩展模块
│   ├── onnx/                       # ONNX 适配
│   └── optim/                      # 优化器适配
├── third_party/                    # 第三方依赖
├── torchnpugen/                    # 代码生成工具
├── examples/                       # 示例代码
└── test/                           # 测试用例
```

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

### torch.nn.ReflectionPad1d

Source: https://github.com/ascend/pytorch/blob/master/docs/api/torch_npu_apis.md

ReflectionPad1d is supported on NPU. It only supports fp16, fp32, int64, and bool data types.

```APIDOC
## torch.nn.ReflectionPad1d

### Description
Pads the input tensor using reflection of the input along the dimension.

### Supported Data Types
- fp16
- fp32
- int64
- bool
```

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

### torch.monitor.Event.__init__

Source: https://github.com/ascend/pytorch/blob/master/docs/api/torch_npu_apis.md

Initializes a new instance of the Event class.

```APIDOC
## torch.monitor.Event.__init__

### Description
Initializes a new instance of the Event class.

### Method
`__init__`
```

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

### Compile torch-npu with Specific Python Version

Source: https://github.com/ascend/pytorch/blob/master/README.md

Execute the build script within the Docker container for a specified Python version. For example, use --python=3.8 for Python 3.8.

```shell
cd /home/pytorch
bash ci/build.sh --python=3.8
```

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

### torch.ao.quantization.prepare

Source: https://github.com/ascend/pytorch/blob/master/docs/api/torch_npu_apis.md

Prepares a PyTorch model for post-training quantization.

```APIDOC
## torch.ao.quantization.prepare

### Description
Prepares a PyTorch model for post-training quantization.

### Method
Not applicable (Python function)

### Endpoint
Not applicable (Python function)

### Parameters
This function takes the PyTorch model as input and prepares it for quantization. Specific parameters are not detailed in the source.

### Request Example
Not applicable (Python function)

### Response
Returns the prepared PyTorch model.

#### Success Response (200)
Not applicable (Python function)

#### Response Example
Not applicable (Python function)
```

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

### Compile torch-npu with Specific PyTorch Version

Source: https://github.com/ascend/pytorch/blob/master/README.md

Build torch-npu against a specific PyTorch version using the --torch flag. The installed PyTorch version must match the specified version.

```shell
bash ci/build.sh --python=3.8 --torch=2.10.0
```

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

### torch.jit.wait

Source: https://github.com/ascend/pytorch/blob/master/docs/api/torch_npu_apis.md

Details on the compatibility and limitations of torch.jit.wait for NPU.

```APIDOC
## torch.jit.wait

### Description
Provides information on the compatibility of `torch.jit.wait` with NPU.

### Compatibility
- NPU Compatibility: Not specified

### Limitations
No specific limitations are documented for NPU usage.
```

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

### Compile torch_npu for a Specific PyTorch Version

Source: https://github.com/ascend/pytorch/blob/master/ci/docker/README.md

Compile torch_npu for a specific PyTorch version. The installed PyTorch version must match the one specified with the --torch argument. Supported versions include 2.10.0, 2.11.0, and 2.12.0.

```Shell
bash ci/build.sh --python=3.9 --torch=2.11.0
```