triton 3.0.0

A language and compiler for custom Deep Learning operations

Stars: 13003, Watchers: 13003, Forks: 1589, Open Issues: 694

The triton-lang/triton repo was created 10 years ago and the last code push was 6 hours ago.
The project is extremely popular with a mindblowing 13003 github stars!

How to Install triton

You can install triton using pip

pip install triton

or add it to a project with poetry

poetry add triton

Package Details

Author

Philippe Tillet

License

None

Homepage

https://github.com/triton-lang/triton/

PyPi:

https://pypi.org/project/triton/

GitHub Repo:

https://github.com/openai/triton

Classifiers

• Software Development/Build Tools

Related Packages

Errors

A list of common triton errors.

Code Examples

Here are some triton code examples and snippets.

GitHub Issues

The triton package has 694 open issues on GitHub

• AxisInfoVisitor for LoadOp constancy calculation
• mac m2 compile
• why BLOCK_SIZE = triton.next_power_of_2(n_cols) is constexpr?
• is the head dimensions of fused attention supported for other values (only 64 now ) ?
• triton.language.sum is incorrect with tensor which is float16 dtype
• pretraining loss blows up with Triton Flash attention
• error: 'triton_gpu.cmpf' op requires the same encoding for all operands
• After build from source, Error occured "device kernel image is invalid"
• Move the default num_warps and num_stage setting from JITFunction to triton.compile
• [OPTIMIZER][BACKEND] significantly cleaner handling of mixed-precision kernels
• Various block_ptr issues
• fp16 tl.sigmoid: "Failed to translate TritonGPU to LLVM IR"
• Incorrect Results with TF32 on Main
• [TUTORIALS] Add invalid memory access analysis tutorial
• Add additive attention mask support for FlashAttention

See more issues on GitHub