why not just compare against a Pytorch reference instead of comparing the kernels to a separate scalar Rust implementation?
(is it for perf reasons because bottleneck becomes verification? Though in that case perhaps it's better to just run a Pytorch reference on the GPUs, which should be good enough. Though, perhaps if you could get this current design to the point where `overlap(CPU, GPU cuda) < GPU cuda + GPU reference` that would be pretty awesome.
Would be cool if you could add coverage, for each GPU, of both rental price as well as purchase price (i.e. buy price := one of the spectrum of completed transactions on ebay, liquidator price, second hand price, new price, MSRP).
Always found the relationship between rental price and purchase price a bit peculiar.
(is it for perf reasons because bottleneck becomes verification? Though in that case perhaps it's better to just run a Pytorch reference on the GPUs, which should be good enough. Though, perhaps if you could get this current design to the point where `overlap(CPU, GPU cuda) < GPU cuda + GPU reference` that would be pretty awesome.