[BUG] compute_patch_embeddings is not using GPU

[dangom13]

When computing embeddings, only CPU is used even though GPUs are available. I'm using the following code snippet:
embeddings_model = foz.load_zoo_model("zero-shot-classification-transformer-torch", name_or_path="google/siglip-so400m-patch14-384")
dataset.compute_patch_embeddings(embeddings_model, detections_field, embeddings_field=embeddings_field)

[Burhan-Q]

Assuming that's a transformers library model, if the hardware is there, you'll need to ensure that PyTorch with CUDA is installed in the same environment as fiftyone or the CPU will get used for computation. Check out the "Getting started locally" page from PyTorch on how to install with CUDA for your OS.

fiftyone/fiftyone/utils/transformers.py

Line 327 in 21468df

d, "device", default="cuda" if torch.cuda.is_available() else "cpu"