No support for multi-GPU

It seems that it's not possible to run models using multiple gpus, e.g. by passing `device_map="auto"` to pipelines.

Is there any way to work around this limitation?

Start by tracing how pipelines are configured and how the `device_map="auto"` argument is handled in the inference toolkit. Reproduce the reported multi-GPU limitation with a model and pipeline, then document the supported behavior or define the changes and validation needed for multi-GPU execution.

Written by the indexing model from the issue text.