Repository navigation
AOT and CubeCL for non-AI software #5418
|
Do cubeCL and Burn have AOT(ahead of time compile) option for things do with JIT, if there is no option for AOT why not? |
Replies: 1 comment 1 reply
|
CubeCL kernels are JIT-compiled, not AOT-compiled. It does have a persistent compilation cache, along with an autotuning cache. But it's not AOT: kernels are still generated when first encountered and then reused on later runs. You can, however, pre-warm those caches and ship them with your application. For known workloads and hardware, this should have similar practical benefits while retaining JIT as a fallback. And yes, CubeCL can absolutely be used for non-AI workloads. It is a general-purpose compute language and runtime. Burn is its main consumer and drives much of its development, so many features are motivated by ML workloads, but CubeCL itself is not limited to this domain. |
|
Thsnk you！ |
CubeCL kernels are JIT-compiled, not AOT-compiled. It does have a persistent compilation cache, along with an autotuning cache. But it's not AOT: kernels are still generated when first encountered and then reused on later runs.
You can, however, pre-warm those caches and ship them with your application. For known workloads and hardware, this should have similar practical benefits while retaining JIT as a fallback.
And yes, CubeCL can absolutely be used for non-AI workloads. It is a general-purpose compute language and runtime. Burn is its main consumer and drives much of its development, so many features are motivated by ML workloads, but CubeCL itself is not limited to this domain.