Large model loading: add link to existing documentation

[LysandreJik]

The documentation for large model loading is in two different places. This adds a link from one to the other, showing the auto device map.

[HuggingFaceDocBuilderDev]

The documentation is not available anymore as the PR was closed or merged.

[sgugger]

Those are actually two different things: the from_pretrained method performs a lot of logic specific to Transformers models that Accelerate (which is model agnostic) does not, so it's unlikely to be useful to the user to go browse that doc. For instance, you can't load an opt model directly in OPTForCausalLM using Accelerate.

So I'd still keep the paragraph about low memory loading.

[LysandreJik]

Hmmm yes, I agree, but this is still linking to from_pretrained, just a part of the documentation that contains more information about the low_cpu_mem_usage. The part I have removed glosses over it quickly, whereas the part I link to has extensive documentation covering both low_cpu_mem_usage and the device_map argument to pass to from_pretrained.

Reading the documentation right now, if we're interested in big models and we open the "Instantiating a big model" page in the toctree, there are no mention of the device_map. This is what this PR aims to fix.

[sgugger]

Ah my mistake, I thought the link was going to the doc of Accelerate 🤦‍♂️