從 Node.js/Python 到 PTX：首個由 AI 生成的深度學習框架

Navigation Menu

Search code, repositories, users, issues, pull requests...

Provide feedback

We read every piece of feedback, and take your input very seriously.

Saved searches

Use saved searches to filter your results more quickly

To see all available qualifiers, see our documentation.

The first fully AI generated DL framework

License

Uh oh!

There was an error while loading. Please reload this page.

NVlabs/vibetensor

Folders and files

Latest commit

History

Repository files navigation

VibeTensor

From Node.js/Python to PTX assembly: a research deep-learning runtime fully generated by AI agents.

Warning

Everything in this repo is generated by AI Agent without human editing. Don't use this project for production purpose.

VibeTensor is a PyTorch-inspired eager runtime with a fresh C++20 core (CPU + CUDA), a torch-style Python overlay, and an experimental Node.js / TypeScript API.

Unlike “thin wrappers”, VibeTensor implements its own tensors/storage, dispatcher, autograd engine, CUDA runtime + caching allocator, and plugin ABI.

Scale (at time of writing): ~218 C++/CUDA source files (~60k non-blank lines) plus ~50k non-blank lines of tests (C++ + Python + JS).

Why this repository exists

VibeTensor started as a system-software experiment: can AI agents generate a coherent deep learning stack spanning Python/JS bindings → a C++ runtime → CUDA allocator/streams → GPU kernels?

The project uses a “least intervention” methodology: high-level architectural prompts, with success measured by whether the system builds, passes tests, and produces correct results.

Performance caveat (“Frankenstein Effect”)

VibeTensor is designed correctness-first and is not performance-competitive with PyTorch today. One failure mode we observed is that components can be individually robust but become globally suboptimal when composed (e.g., global serialization in the hot path).

Treat VibeTensor as a research prototype.

Architecture (macro view)

At a high level:

Module Architecture Diagrams

For more detailed architecture of individual modules, see the diagrams in docs/images/:

Project status

VibeTensor is an active research project; APIs and behavior may change without notice.

Current code covers:

Platforms & requirements

Install & build

Editable dev install (recommended)

From the repository root:

This drives CMake via scikit-build-core (build dir: build-py/), building:

Defaults are configured in pyproject.toml; override via -Ccmake.define.* (for example, -Ccmake.define.VBT_BUILD_NODE=OFF).
If CMake can't find node_api.h, set NODE_INCLUDE_DIR=/path/to/include/node (or pass -Ccmake.define.NODEJS_INCLUDE_DIR=...).

Manual CMake build

Wheels (Release)

Quickstart

Python: torch-style overlay

The main user-facing API is vibetensor.torch, a lightweight torch-like overlay built on top of the C++ core.

CUDA helpers (streams, events, memory stats) live under vibetensor.torch.cuda:

For lower-level access (bindings/tests), import the native extension directly:

Node.js / TypeScript (experimental)

The Node addon and JS/TS overlay live under js/vibetensor/.

After an editable install has produced js/vibetensor/vbt_napi.node, you can run the JS tests locally:

A minimal usage example (from js/vibetensor after npm run build):

The JS overlay is fully async: heavy work (CUDA ops, H2D/D2H copies, waits) is scheduled via napi_async_work and does not block the Node event loop.

Note: JS Tensor objects are currently CPU-only; CUDA is exposed via cuda.* (streams/events), explicit h2d/d2h, and DLPack import/export.

Dynamic operators, plugins, and Triton

VibeTensor exposes several extension points:

Simple Python override:

Loading a C/CUDA plugin:

Testing

After an editable install, typical test commands are:

Note: Python tests require numpy; some tests also require torch and/or triton and will skip if unavailable.

C++ tests (CTest)

Python tests (pytest)

Node / JS overlay tests

The Node step is treated as best-effort in CI: when js/vibetensor/vbt_napi.node is missing, JS tests are skipped with a success status.

API parity gate

Repository layout

Credit

License

VibeTensor is licensed under the Apache License, Version 2.0. See pyproject.toml and the repository's license files for details. Vendored third-party code remains under its original licenses.

About

The first fully AI generated DL framework

Topics

Resources

License

Contributing

Uh oh!

There was an error while loading. Please reload this page.

Stars

Watchers

Forks

Releases

Packages

  0

Languages

Footer

Footer navigation