1. To enable the Ascend AI Processor to use the capabilities of the PyTorch framework, the native PyTorch framework needs to be adapted at the device layer. The APIs related to the CPU and CUDA need to be replaced for external presentation. During network porting, some device-related APIs need to be replaced with the APIs related to the Ascend AI Processor. The following table lists the supported device-related APIs.
   Table 6-2 Device-related APIs

   Original PyTorch API	API Adapted to the Ascend AI Processor	Description
   torch.cuda.is_available()	torch.npu.is_available()	Checks whether the device is available in the current environment (not the final result).
   torch.cuda.current_device()	torch.npu.current_device()	Obtains the device in use.
   torch.cuda.device_count()	torch.npu.device_count()	Obtains the number of devices in the current environment.
   torch.cuda.set_device()	torch.npu.set_device()	Sets the device in use.
   torch.tensor([1,2,3]).is_cuda	torch.tensor([1,2,3]).is_npu	Checks whether a tensor is in the format on the CUDA or NPU device.
   torch.tensor([1,2,3]).cuda()	torch.tensor([1,2,3]).npu()	Converts a tensor to the format on the CUDA or NPU device.
   torch.tensor([1,2,3]).to("cuda")	torch.tensor([1,2,3]).to('npu')	Converts a tensor to the format on the CUDA or NPU device.
   torch.cuda.synchronize()	torch.npu.synchronize()	Waits until the event is complete.
   torch.cuda.device	torch.npu.device	Generates a device class, which can be used to perform device-related operations.
   torch.cuda.Stream(device)	torch.npu.Stream(device)	Generates a stream object.
   torch.cuda.stream(Stream)	torch.npu.stream(Stream)	Mainly used for scope restriction.
   torch.cuda.current_stream()	torch.npu.current_stream()	Obtains the current stream.
   torch.cuda.default_stream()	torch.npu.default_stream()	Obtains the default stream.
   device = torch.device("cuda:0")	device = torch.device("npu:0")	Specifies a device.
   torch.autograd.profiler.profile (use_cuda=True)	torch.autograd.profiler.profile (use_npu=True)	Specifies that CUDA/NPU is used during profiler execution.
   torch.cuda.Event()	torch.npu.Event()	Returns events on a device.

2. When building or porting a network, you need to create tensors of specified data types. The following table lists the tensors created on the Ascend AI Processor.
   Table 6-3 Tensor-related APIs

   GPU tensor	API Adapted to the Ascend AI Processor
   torch.tensor([1,2,3],dtype=torch.long,device='cuda')	torch.tensor([1,2,3],dtype=torch.long,device='npu')
   torch.tensor([1,2,3],dtype=torch.int,device='cuda')	torch.tensor([1,2,3],dtype=torch.int,device='npu')
   torch.tensor([1,2,3],dtype=torch.half,device='cuda')	torch.tensor([1,2,3],dtype=torch.half,device='npu')
   torch.tensor([1,2,3],dtype=torch.float,device='cuda')	torch.tensor([1,2,3],dtype=torch.float,device='npu')
   torch.tensor([1,2,3],dtype=torch.bool,device='cuda')	torch.tensor([1,2,3],dtype=torch.bool,device='npu')
   torch.cuda.BoolTensor([1,2,3])	torch.npu.BoolTensor([1,2,3])
   torch.cuda.FloatTensor([1,2,3])	torch.npu.FloatTensor([1,2,3])
   torch.cuda.IntTensor([1,2,3])	torch.npu.IntTensor([1,2,3])
   torch.cuda.LongTensor([1,2,3])	torch.npu.LongTensor([1,2,3])
   torch.cuda.HalfTensor([1,2,3])	torch.npu.HalfTensor([1,2,3])

For more APIs, see the PyTorch API Support List.