Wondering whether most major inference libraries support storage-direct or if the listed providers are cheaping out on storage latency. Several seconds to load

Wondering whether most major inference libraries support storage-direct or if the listed providers are cheaping out on storage latency. Several seconds to load a 100MB model when pcie 4 is 256Gb/s I'd have expected an order of magnitude less - at least from my experience with gpudirect and real-time processing of up to 200Gb/s network streams with datacenter gpus (not performing inference but I'm talking about actual streaming to gpu memory here). PCIe-5 and H100 (the only one in the line-up that has PCIe-5...) should also improve on that.