https://docarray.jina.ai/advanced/torch-support/

The bold part is the description that I think is problematic

There is a very common pattern in the deep learning engineering: loading big data, mapping it via some function for preprocessing on GPU, and batching it to GPU for intensive deep learning stuff.

There are many pitfalls in this pattern when not implemented correctly, to name a few:

• data may not fit into memory;
• mapping via CPU only utilizes a single-core;
• data-draining problem: GPU is not fully utilized as data is blocked by the slow CPU preprocessing step.

Combined with context inference, the correct description should be CPU rather than GPU.
It should be revised as follows:

loading big data, mapping it via some function for preprocessing on CPU,