Providing AI agent SKILLS for using IRON, as opposed to working with it for contributing, may help increase adoption of IRON and the boys in the Ryzen AI series APUs. They should cover all available operators and their input and output constraints as well as how to use them. This should include operators not mentioned in the README such as Repeat that might be fundamentally required for high level composition to perform inference using IRON. Additionally, it may be helpful to split them by supported hardware so that the agent consuming the skill has access to the grossly undocumented information it needs about the hardware to avoid compile errors from hardware constraints.
I tried to have Claude right a SKILL specifically for using IRON on aie2p, but it has flaws and things are missing even after Claude ate through all my quota to do it. This resulted in having to point less "intelligent" models to the source code and kernels in my virtual environment so it could figure out how to use it. I then had to fight the agent extensively to prevent it from alternating the installed IRON package and stop it from attempting to do so (fun times 😅).
I feel like the npus on these systems are grossly underutilized when they are extremely capable and hope to see wider adoption from the community. The greater the adoption of IRON the more likely it is for users to contribute kernels and thus create a cycle of growth and adoption.
Including how to create a custom kernel external to the iron repo may be useful too to close any gaps encountered, which can then be contributed later.
Providing AI agent SKILLS for using IRON, as opposed to working with it for contributing, may help increase adoption of IRON and the boys in the Ryzen AI series APUs. They should cover all available operators and their input and output constraints as well as how to use them. This should include operators not mentioned in the README such as Repeat that might be fundamentally required for high level composition to perform inference using IRON. Additionally, it may be helpful to split them by supported hardware so that the agent consuming the skill has access to the grossly undocumented information it needs about the hardware to avoid compile errors from hardware constraints.
I tried to have Claude right a SKILL specifically for using IRON on aie2p, but it has flaws and things are missing even after Claude ate through all my quota to do it. This resulted in having to point less "intelligent" models to the source code and kernels in my virtual environment so it could figure out how to use it. I then had to fight the agent extensively to prevent it from alternating the installed IRON package and stop it from attempting to do so (fun times 😅).
I feel like the npus on these systems are grossly underutilized when they are extremely capable and hope to see wider adoption from the community. The greater the adoption of IRON the more likely it is for users to contribute kernels and thus create a cycle of growth and adoption.
Including how to create a custom kernel external to the iron repo may be useful too to close any gaps encountered, which can then be contributed later.