A learner-friendly, PyTorch-compatible deep learning framework designed to demystify neural network internals and facilitate custom hardware experimentation.
Select your preferences and run the install command
pip install tensorplay --index-url https://download.tensorplay.cn/whl/cu130A collection of libraries designed to be used together or independently.
The core calculation engine. Provides a clean, readable implementation of memory management and basic tensor kernels.
Design Philosophy →The explicit autograd layer. Designed for those who want to understand or modify how computational graphs are built.
Design Philosophy →The optimization playground. Experiment with operator fusion and static graph capture in a simplified environment.
Design Philosophy →The modular business layer. High-level components (Linear, Conv2d) that serve as blueprints for your own network layers.
Design Philosophy →Stay updated with our latest blog posts and research highlights.
深入探讨 TensorPlay “解耦优先”的设计哲学。通过将框架拆分为 P10、TPX、Stax 和 NN 四个独立且职责单一的库，我们实现了灵活性与高度可定制性的完美平衡。
揭秘 TensorPlay 如何在底层实现跨硬件的算子调度。通过 Dispatcher 机制与 TensorImpl 模式，P10 能够以极低的延迟将计算任务分发至最合适的硬件后端。
传统的 autograd 往往深度嵌入计算逻辑。TPX 采用了一种全新的“透明扩展”模式，在完全不侵入 P10 计算逻辑的前提下，实现了灵活的动态图追踪与梯度回传。