how to optimize setfit inference

[geraldstanje]

hi,

im currently investigating what the options we have to optimize setfit inference and have a few questions about it:

• gpu:
  • torch compile: https://huggingface.co/docs/transformers/en/perf_torch_compile
    is the following the only way to use setfit with torch.compile?

model.model_body[0].auto_model = torch.compile(model.model_body[0].auto_model)

info above was provided by Tom Aarsen.

does torch.compile also work for cpu? edit: looks like it should work for cpu too...

https://pytorch.org/docs/stable/generated/torch.compile.html
does torch compile change anything about the accuracy of the model inference?

i see different modes here:
Can be either “default”, “reduce-overhead”, “max-autotune” or “max-autotune-no-cudagraphs” ... so far reduce-overhead gives best results....

• cpu:
  what are the options to optimize cpu inference?
  • BetterTransformer: https://huggingface.co/docs/transformers/en/perf_infer_cpu
    is BetterTransformer really not available for setFit? i dont see setFit in this list: https://huggingface.co/docs/optimum/bettertransformer/overview#supported-models

are there any other resources to speedup setfit model inference? where can you run a setFit model except torchServe?

Thanks,
Gerald