Huawei Debuts CANN PyPTO Framework with Day-Zero Support for Ant's Ling-3.0-Flash
Huawei has announced day-zero support for Ant Group's Ling-3.0-flash model on its Ascend A2 and A3 hardware, while debuting CANN PyPTO, a Python-based operator programming framework. The framework features CANNBot, a multi-agent system that automates the entire operator development process across seven stages. This release simplifies and accelerates the deployment of complex AI models on Huawei's Ascend chips by allowing developers to write custom operators in Python instead of low-level hardware code. It demonstrates Huawei's commitment to building a robust software ecosystem that can rapidly adapt to new open-source models like Ling-3.0-flash. CANN PyPTO uses a Parallel Tensor/Tile Operation (PTO) virtual instruction set to automate tiling, data movement, and code generation. Running the Ling-3.0-flash model requires a minimum hardware configuration of either eight Ascend A2 cards or four Ascend A3 cards.
## BACKGROUND
CANN (Compute Architecture for Neural Networks) is Huawei's proprietary chip-enabling platform for Ascend AI processors, serving a similar role to NVIDIA's CUDA. Operators are the basic mathematical building blocks (like matrix multiplication) used to construct neural networks, and optimizing them is crucial for maximizing hardware performance.