I want insights and experiences related to exploring the TOSA specification

TOSA
#1

Hey everyone,

I have been diving into the TOSA lately & I am really excited about its potential for enhancing machine learning workflows.

For those who have already worked with TOSA what aspects do you find most beneficial?? Are there any specific use cases or applications where you have seen it shine? I am interested in how it can improve model and performance across different platforms in AI applications.

As well, I found these resources when doing research on this; TOSA Specification 0.22.0 released & if anyone have any resources, tutorials or personal experiences please share with me, It would be greatly appreciated!!

Thank you……. :slight_smile:

#2

Hi Noah,

Thanks for looking into TOSA and posting. There are a lot of ways that TOSA can be useful to you, depending on what you’re looking for. One of the primary reasons we developed TOSA was to provide a consistent managed set of tensor operators that can be used for everything from designing custom hardware to defining a smaller set of software kernels that need to be optimized.

The specification is a good place to start, we tried to list a lot of the principles we used when defining TOSA. The 0.22 version is pretty old, you can find a more recent version here: TOSA 0.80.0 specification.

One area some people find useful is the TOSA MLIR dialect: Tensor Operator Set Architecture (TOSA) Dialect - MLIR

Hopefully that gives you a few places to look. Feel free to ask more questions if there’s anything I can help explain.

Eric