Python frontendtensor(x)
User code creates tensors and describes computation; devices stay invisible.
eagerRunning
SEE THE SYSTEM. CHANGE IT.
Trace every tensor to its kernel. A transparent framework for learning, research, and hardware experiments.
Python / C++ coreCUDA · ROCm backends
No hidden compute path. Read, verify, and recompose every forward pass, gradient, and kernel execution.
01Forward pass
02Autograd
03Kernel execution
Watch a network learn.
Training in the open, no black box: weights, gradients, and the decision boundary unfold step by step.
Play to start · click the canvas to restart
Nothing hidden
How gradients update and how the boundary forms — every step unfolds in front of you.
Weights are the lines
Line thickness is the weight; the whole network reads at a glance.
A boundary that breathes
The heatmap is the network's current judgment, settling onto the data as it learns.
Follow the tensor.
From Python to a native graph, then gradients and a kernel. Every boundary stays readable.
FROM OPS
Every op call. One pipeline.
TO RUNTIME
One runtime takes all of them.
TensorPlay
tensor(x)Python frontend · accepted
tp.trace(fn)TPX Graph · accepted
loss.backward()Autograd · accepted
graph.compile()Compiler passes · accepted
kernel.launch()CUDA backend · accepted
kernel.launch()ROCm backend · accepted
kernel.run()CPU backend · accepted
dtensor.shard()Distributed runtime · accepted
optimizer.step()Optimizer · accepted
tp.profile()Profiler · accepted
state.save()Checkpoint · accepted
graph.export()Serialization · accepted
Clarity beats magic.
Pure and transparent
Trace every operation from Python to the C++ core. Peel back each abstraction and keep the full path readable.
DIY acceleration
Readable CPU and CUDA backends leave room for custom kernels, new hardware, and nanobind experiments.
Modular autograd
TPX decouples computation from gradients, making dynamic graphs and new differentiation modes easier to modify.
Research ready
Prototype layers, optimizers, storage formats, operator fusion, and static graph capture with less boilerplate.
Four layers. No black box.
P10
Execute
The pure tensor engine for memory, kernels, dispatch, and hardware backends.
TPX
Trace
The explicit autograd layer for dynamic graphs, gradients, and backward execution.
Stax
Accelerate
The static graph and JIT playground for fusion, capture, and compute optimization.
NN
Build
The high-level neural network layer, from Linear and Conv2d to custom modules.
Run it. Then open it.
Choose your system, install method, and compute platform. We generate the command for you.
PYTHON 3.10+CUDA 12.4 · 12.6 · 13.0ROCM 7.2
Platform
Install method
LanguagePython
Compute platform
Run this commandbash
python3 -m pip install tensorplay --upgradeDispatches from inside the system.
Keep going.
Open the black box with us.
Students, developers, educators, and researchers are welcome. Learn the system, contribute code, discuss architecture, or follow project updates.