Will it run?

Pick a model and a workload. We size memory (weights, KV cache, optimizer state), shard it across chips, and check whether the compiler stack actually supports it.