Why is it taking so much RAM?

[fernandocamargoai]

I'm packing my model for inference and, if I have a GPU available, I transfer my model to it and hybridize it with the following code:

predictor.ctx = self.ctx
predictor.prediction_net.collect_params().reset_ctx(self.ctx)

data_entry, _ = self._query_and_preprocess(self.sample_request.jsonData)
dataset = self._prepare_dataset([data_entry])

inference_data_loader = InferenceDataLoader(
    dataset,
    transform=predictor.input_transform,
    batch_size=predictor.batch_size,
    stack_fn=functools.partial(batchify, ctx=predictor.ctx, dtype=predictor.dtype),
)

batch = next(iter(inference_data_loader))

predictor.hybridize(batch)

When my this code is finished, my model occupies 5GB of RAM and 600MB of VRAM. The VRAM is pretty normal, but I don't understand why this much RAM is needed, especially because my model is in GPU. If I don't transfer it to GPU (removing the first two lines of code), it takes only 550MB of RAM. If I only transfer it to GPU (lefting only the first two lines of code), it takes 1.8GB of RAM. So, this weird behavior only happens when using GPU. Am I doing something wrong?