HAT use cases in AI

[Paul Sandoz]

Hi Andrii,

I think there is plenty of room to explore many forms of AI development, including using ONNX, and more broadly code reflection can apply to other many other non-AI use cases.

I would like to explore the expression of ONNX functions and models using a subset of Java code, much like ONNX Script [1] does for Python.

The GPU/HAT use-case explores a parallel computing model, applicable HPC of which AI is a subset, and targeting GPU execution environments.

Not all AI is deep learning.

Very recently Oracle open sources Sandwood [2], a language, compiler, and runtime for JVM based probabilistic models. Code reflection can be used to express the probabilistic model in Java code rather than in its own domain specific language. HAT might also be used for executing the probabilistic models on the GPU. 

Another example is MSET2 [3], we can implement on the GPU using HAT, or on the CPU using vectorization where we could use code reflection to vectorize lanewise expressions.

Hopefully you can see from those few examples there are a broad set of use cases where HAT and/or just code reflection may be applicable.

Paul.


[1] https://github.com/microsoft/onnxscript
[2] https://github.com/oracle/sandwood
[3] https://docs.oracle.com/en-us/iaas/Content/anomaly/using/kernels.htm


> On Aug 8, 2024, at 8:54 AM, Andrii Lomakin <andrii0lomakin at gmail.com> wrote:
> 
> Good day.
> 
> Probably it is a very preliminary question to ask.
> 
> But what are the benefits of using Bablylon/HAT the project's creators
> can see, in the context of AI development, if we take into account
> that there are already big players such as ONX Runtime?
> 
> I am asking to understand the vision of the project creators without
> any intention of undermining the project's value.