'%20stroke-opacity='0'%20stroke-width='0.700000'%20style=''/%3e%3cpath%20d='M37.0859%2021.1797C39.5239%2023.7207%2040.8945%2027.1738%2040.8945%2030.7754C40.8945%2034.3774%2039.5234%2037.8311%2037.0847%2040.3721C34.6816%2042.9072%2031.4111%2044.333%2027.9998%2044.333C24.5881%2044.333%2021.3179%2042.9072%2018.9146%2040.3721C16.4761%2037.8311%2015.105%2034.3774%2015.105%2030.7754C15.105%2027.1733%2016.4761%2023.7197%2018.9146%2021.1787L27.9998%2011.6665L27.9998%2016.4121L27.1519%2017.2891L21.095%2023.6309C17.3201%2027.5854%2017.3213%2033.9951%2021.0977%2037.9482C24.874%2041.9019%2030.9961%2041.9019%2034.7727%2037.9482C38.5491%2033.9951%2038.5503%2027.5854%2034.7751%2023.6309L37.0859%2021.1797ZM37.0859%2021.1797L37.0869%2021.1787L37.0847%2021.1787L37.0859%2021.1797ZM31.0674%2019.7607C31.3735%2020.3086%2031.9353%2020.6436%2032.5413%2020.6396C33.4778%2020.6338%2034.2327%2019.834%2034.2271%2018.8535C34.2214%2017.873%2033.4575%2017.0825%2032.521%2017.0884C31.915%2017.0923%2031.3572%2017.4341%2031.0574%2017.9854C30.7576%2018.5366%2030.7615%2019.2134%2031.0674%2019.7607Z'%20clip-rule='evenodd'%20fill='%23FFFFFF'%20fill-opacity='1.000000'%20fill-rule='evenodd'/%3e%3c/svg%3e)
PyTorch Adaptation
Community-driven, Ascend-adapted PyTorch supports mainstream third-party ecosystems of PyTorch.
Adaptation to Ascend AI Processors
Built upon open-source PyTorch, it adapts to Ascend AI Processors and offers native Python interfaces.
Core Framework Functions
Provides PyTorch dynamic graphs, automatic differential, profiling, optimizers, etc.
Custom Operator Development
Allows adding custom operators within the PyTorch framework.
Distributed Training
Supports native distributed data parallelism, including collective communication primitives, such as Broadcast and AllReduce, for single-server multi-device and multi-server multi-device scenarios.
Mixed Precision
Facilitates O1 and O2 mixed precision along with static and dynamic Loss Scale algorithms, allowing precision configuration of operators via blocklists and trustlists. Download and install the APEX package for more details.
Model Inference
Outputs standard ONNX models and converts them into offline inference models via the offline conversion tool.
Development Scenarios
Model Script Adaptation
Automatically or with minimal manual effort, adapt model scripts from other AI platforms to Ascend APIs for Ascend AI Processor-powered model training.
Model Training
Develop and train network models based on native PyTorch APIs and Ascend APIs, while using the Ascend tuning tool to tune training accuracy and performance, to deliver high-precision and high-performance AI algorithms.
Operator Development
Develop Ascend AI Processor-powered custom operators using the Ascend development toolchain to deeply optimize network models.
Learning Resources
Getting Started
Get started with the PyTorch framework to set up an environment for migration and training.
Migration and Training
Comprehensively understand the PyTorch framework to complete model migration and training.
Training and Tuning
Evaluate accuracy and performance across all scenarios from multiple perspectives.