From custom ops to native graphs
Follow custom_op, Triton, TVM, and Stax native lowering through one executable and verifiable kernel path.
Read articleBlog and news
Design notes from dispatch and hardware abstraction to automatic differentiation.
The TensorPlay compiler roadmap
From graph capture, PassManager, and shape guards to AOT autograd and Stax native execution.
Read articleWhy TensorPlay uses four core libraries
How P10, TPX, Stax, and NN use one-way dependencies to separate compute, gradients, graph optimization, and application logic.
Read articleInside the P10 Dispatcher
From TensorImpl to DispatchKey, trace one tensor interface across CPU, CUDA, and custom hardware.
Read articleHow TPX separates computation and gradients
Build a complete dynamic graph through composition, on-demand tracing, and GradFn scheduling without invading the compute core.
Read article