fix(trainer): Avoid moving model with device_map (#41032)
* fix(trainer): Avoid moving model with device_map
When a model is loaded with `device_map="auto"` and is too large to fit on a single GPU, `accelerate` will offload some layers to the CPU or disk. The `Trainer` would previously attempt to move the entire model to the specified device, causing a `RuntimeError` because a model dispatched with `accelerate` hooks cannot be moved.
This commit fixes the issue by adding a check in `_move_model_to_device` to see if the model has an `hf_device_map` attribute. If it does, the device placement is assumed to be handled by `accelerate`, and the `model.to(device)` call is skipped.
A regression test is added to ensure the `Trainer` can be initialized with a model that has a `hf_device_map` that simulates offloading without raising an error.
* Added the logger warning for the move model
---------
Co-authored-by: google-labs-jules[bot] <161369871+google-labs-jules[bot]@users.noreply.github.com>
