What's New in 7.0.RC1

Compared with MindStudio 6.0.0, MindStudio 7.0.RC1 has the following changes:

New Features

Document	New Features
Operator Development Tools	To verify the functions of the tools under the PyTorch framework, install the PyTorch and Ascend PyTorch Adapter in advance. To use the simulation pipeline function, generate simulation dump data in advance.
Analysis and Migration	X2MindSpore MindSpore supports two running modes, Graph and PyNative. Due to Python syntax restrictions, the Graph mode is switched to the PyNative mode, which delivers lower performance. To avoid the restriction that tensors cannot be created during data processing in MindSpore, you can set the running mode to synchronous operator delivery during PyTorch conversion. As a result, the training performance may deteriorate. In this case, you can remove pynative_synchronize=True from context.set_context and use asynchronous operator delivery to improve the performance. If an error is reported, check the data processing code, remove the tensor creation part, and use NumPy NDarray. Currently, the ported TensorFlow models cannot run on multiple devices. Cloud-based porting is not supported for TensorFlow 1.x, and static graph porting is not supported for TensorFlow 1.x/2.x. PyTorch GPU2Ascend The migration tool converts scripts by the adaptation rules, significantly accelerating script migration and reducing development workload. The original scripts must be executed in the GPU environment and based on Python 3.7 or later. The execution logic after conversion must be the same as that before conversion. This tool supports the conversion of training scripts of PyTorch 1.11.0, 2.0.1, and 2.1.0.

Document

New Features

Operator Development Tools

To verify the functions of the tools under the PyTorch framework, install the PyTorch and Ascend PyTorch Adapter in advance.
To use the simulation pipeline function, generate simulation dump data in advance.

Analysis and Migration

X2MindSpore

MindSpore supports two running modes, Graph and PyNative. Due to Python syntax restrictions, the Graph mode is switched to the PyNative mode, which delivers lower performance.
To avoid the restriction that tensors cannot be created during data processing in MindSpore, you can set the running mode to synchronous operator delivery during PyTorch conversion. As a result, the training performance may deteriorate. In this case, you can remove pynative_synchronize=True from context.set_context and use asynchronous operator delivery to improve the performance. If an error is reported, check the data processing code, remove the tensor creation part, and use NumPy NDarray.
Currently, the ported TensorFlow models cannot run on multiple devices. Cloud-based porting is not supported for TensorFlow 1.x, and static graph porting is not supported for TensorFlow 1.x/2.x.

PyTorch GPU2Ascend

The migration tool converts scripts by the adaptation rules, significantly accelerating script migration and reducing development workload.
The original scripts must be executed in the GPU environment and based on Python 3.7 or later.
The execution logic after conversion must be the same as that before conversion.
This tool supports the conversion of training scripts of PyTorch 1.11.0, 2.0.1, and 2.1.0.

Removed Features

Document	Removed Features
Performance Tuning Tool	Using the CANN Profiling API to collect PyTorch profile data is deleted from the performance tuning tool. The timeline and summary directories are deleted from the performance data parsing result of the performance tuning tool.

Parent topic: Feature Change