HLIR Robustness

[jafioti]

Luminal represents all models in HLIR before entering the search space. This means if an operation isn't representable in HLIR, we can't do it. Therefore its very important HLIR is robust enough to cover all usecases.

Common benchmarks are:

• Does it cover StableHLO?
• Is it enough to specify common LLMs, VLMs, CNNs, etc.
• Does it cover primtorch / Aten?

These will need to be answered over the coming months so that the large investment of building backends is done on a solid HLIR spec.