High inference lantency

[Altimis]

I fine-tuned Donut on custom data, but inference takes 15 seconds on CPU (8 cores). I realized the generate() function is the one that takes too long. Are there any improvements that could be made to reduce this latency please ? Thanks