Investigate support for TIRx (Apache TVM) kernels

[moon-bot-app]

Tracking issue to investigate supporting TIRx from Apache TVM.

Context

TIRx is gaining traction, with kernel authors reporting strong performance numbers (reportedly beating quack/CuteDSL in some cases). It would be valuable to support it in kernels.

Key consideration

Unlike Triton/Helion, TIRx explicitly separates compilation from runtime, so it behaves more like an ahead-of-time (AOT) framework than a JIT DSL. This means integration is likely more involved than simply adding a Python dependency (as we did for Helion).

Proposed direction / open questions

• Update the builder to compile TIRx into a cubin — or possibly emit CUDA C source (which may be preferable) — and feed it through the existing AOT build pipeline.
• Investigate feasibility in more detail.
• Decide whether adding this DSL path to the builder is something we want to maintain long term.

---

Requested by Sayak Paul - Slack thread - Agent trace

[sayakpaul]

Cc: @kashif

[sayakpaul]

Assigned to @drbh as he's working on it.

[danieldk]

We had some discussion about DSL during the last sync. My point of view is:

• We should only add (more) DSLs if there are very performant kernels that downstream users (e.g. transformers) need. Currently there are a lot of new DSLs and it probably makes most sense for us to see which ones get picked up long-term.
• We are working on a kernel-kernel dependency mechanism. Some DSLs may be packagable as kernels. This would then be preferred, since the DSL developers can then maintain it on the hub and it avoids adding more Python dependencies (which is bad for kernels, because it requires user interaction before they can use a kernel, introduces versioning conflicts, etc.).