I have Questions Around Operand Behavior & Real-World Scenarios in TOSA

Hi All,

I have been going through the TOSA & I am a bit stuck on understanding how operands behave, especially when it comes to broadcasting & certain element-wise operations. The spec gives a general idea but when it comes to actual implementation or integration with frameworks, things start to feel a bit fuzzy.

Has anyone run into challenges where TOSA’s constraints did not align well with real-world use cases such as in mobile deployments or with quantized models? I appreciate if someone could share practical examples or edge cases where they had to rethink how they were using certain ops.

Also, while I am diving into this, I am exploring CSM Training on the side, just to strengthen my overall grasp of workflows & team coordination in ML projects—so if there is any crossover insight there. Also I have check this I have Questions Around Operand Behavior & Real-World Scenarios in TOSA still need advice.

I want to hearing your thoughts.

Thank you.:slight_smile:

For operator implementation, the pseudocode is meant to give additional information about how each operator works. Without specifics, it’s hard to give much more advice.

To understand broadcasting, you can look at how NumPy broadcasting works: Broadcasting — NumPy v2.3 Manual. TOSA broadcast is more restrictive than the NumPy version, but the conceptrs are quite similar.