Inaccuracies in PyTorch backend documentation #8853

triton-inference-server/pytorch_backend#117 introduced a Python-based backend for PyTorch 2.0 (model.py). If you look at the state of the README in that PR, there is a section "Customizations" that only applies to model.py, not libtorch or other PyTorch backends.

More recently, triton-inference-server/pytorch_backend#160 and triton-inference-server/pytorch_backend#164 changed the documentation. All mentions of model.py are removed (except for one leftover at the very bottom of the README). The "Customizations" section for model.py is now just part of the overall "Configuration" section, but this is incorrect. The actual thread control parameters for the libtorch backend are INTER_OP_THREAD_COUNT and INTRA_OP_THREAD_COUNT, as listed earlier in the "Parameters" section.

There is also a broken link related to this topic:

[!TIP] Refer to CPU Threading TorchScript on how to set this parameter properly.

The page https://pytorch.org/docs/stable/notes/cpu_threading_torchscript_inference.html now just says "TorchScript is deprecated, please use torch.export instead."

This is unfortunate, because there seems to be no adequate replacement (I can only find https://docs.pytorch.org/docs/2.12/notes/multiprocessing.html#avoid-cpu-oversubscription, which is not nearly as thorough). Perhaps linking to the previous version of the page, https://docs.pytorch.org/docs/2.8/notes/cpu_threading_torchscript_inference.html, is the right choice.

Tagging experts here: @whoisj @kthui @yinggeh

(I am not sure of the right way to fix this, because it depends on the status and outlook for the Python backend.)