PyTorch in Production: Boosting LLM Training and...

Sponsored Session: Lightning Talk: PyTorch in Production: Boosting LLM Training and Inferencing on Ascend NPU - Fengchun Hua, Huawei In scenarios of diverse computing power, PyTorch faces significant challenges in stability, ease of use, performance, and quality. This session will showcase the best practices of PyTorch and its ecosystem components in heterogeneous computing by integrating Huawei’s community contributions and the actual application scenarios of the Ascend platform. Through technical and architectural improvements optimized for the Ascend platform in PyTorch, we have significantly enhanced the computational performance of training and inference, bringing it up to production-grade standards. We are committed to promoting the application and implementation of PyTorch ecosystem components in China, such as the applications of PyTorch ecosystem projects like DeepSpeed, vLLM, torchtune, and torchtitan on Ascend. Finally, we will share real user cases of training and distributed inference, demonstrate the best practices of the Ascend platform, and help community users address challenges in various computing scenarios.