Home

Read the whole stack. A learner-first, DIY-ready framework for tensors, kernels, and custom hardware.

TensorPlay is a learner-friendly, DIY-ready deep learning framework designed to make neural-network internals visible and custom hardware experimentation practical. The whole stack — Python API, C++ core, CUDA kernels, compiler — is engineered to be read: clean implementations, explicit computation graphs, and no black box between your model and the hardware.

Wiki directory

Page	Contents
Getting Started	Install (PyPI / CUDA / nightly / from source) and run the first examples
Architecture	The four pillars — tensorplay, p10, tpx, stax — and where to dig deeper
FAQ	Out-of-memory fixes, build troubleshooting, testing, release channels

Quick links

• Documentation
• Tutorials — linear regression from scratch, MNIST CNN, custom datasets, model saving/loading
• Source code
• Issue tracker
• Discussions
• Releases
• White paper — the whole stack with path:line citations
• Blog series — one pillar per post
• Benchmarks
• Contributing guide
• Security policy

Architecture

• tensorplay — Python API for tensors, autograd, neural networks, optimizers, data loading, and serialization.
• p10 — C++ core engine for tensor storage, memory management, and CPU/CUDA kernels.
• tpx — Explicit autograd and computation-graph execution layer.
• stax — JIT compiler playground for graph capture, lowering, and operator fusion.
• tensorplay.nn — Neural-network building blocks such as modules, layers, activations, and losses.
• tensorplay.optim — Optimizers and learning-rate scheduling.

See the Architecture page for the design rationale, benchmark highlights, and deep-dive material.

Development and releases

The main branch is the active development line. Release candidates use versions such as X.Y.0rcN; stable releases use X.Y.0; nightly builds use date-based development versions. The package version lives in version.txt, and the wheel matrix (platforms, CUDA variants, the ROCm channel) is documented in RELEASE.md.

For contribution workflow, issue triage, release notes, and CI conventions, see the repository documentation and templates linked above.

Community

• Discord — questions, ideas, showcase
• Issue tracker — bug reports, feature requests, RFCs
• Email: feedback@tensorplay.cn

Bug reports, feature requests, performance regressions, and configuration questions should use the appropriate issue template. Security vulnerabilities should be reported privately through GitHub's security reporting flow.

License

TensorPlay is licensed under the Apache 2.0 License.