Parallelization with single model loaded in GPU/VRAM

[flariut]

Hi, I'm running a project trying to parallelize tasks with whisper, I'm using CUDA and loading the model in VRAM. Is there a way to parallelize tasks with a single large-v3 model loaded in memory (~12GB VRAM) instead of loading the model for each task I need concurrency? My method right now is using python's multiprocessing for loading the model twice in memory for having at least 2x, I've tried threading and using the same model for multiple transcriptions but an exception of tensor's weights is raised. Thanks in advance!

[flariut]

Is this possible in this python implementation, or not? it may depend on the implementation? Thanks.