AttributeError: module 'mirage' has no attribute 'generate_triton_program'

[2niuhe]

It seems that generate_triton_program is not supported in the current version.

In [7]: mi.generate_triton_program
---------------------------------------------------------------------------
AttributeError                            Traceback (most recent call last)
AttributeError                            Traceback (most recent call last)
<ipython-input-7-9ef5e8a2b90e> in <cell line: 1>()
----> 1 mi.generate_triton_program

AttributeError: module 'mirage' has no attribute 'generate_triton_program'

[jiazhihao]

The transpiler currently supports CUDA code gen. The Mirage team is discussing a long term plan for Triton code gen. Is there any reason you may prefer Triton code gen to CUDA code gen?

[2niuhe]

I prefer Triton code generation for several reasons. Triton provides a higher-level abstraction that makes it easier to support compilers from various hardware vendors, not just Nvidia. This allows for greater compatibility and enables the use of different manufacturers' Triton compiler ecosystems with minimal modifications.

Additionally, Triton code can be easily integrated with PyTorch, making it more accessible for those who may not be proficient in C++. This flexibility and ease of use are significant advantages for my projects.

Thank you for considering this in your discussions!