JavierS - 

The dbutils serialization error occurs in your code because dbutils is only available on the Databricks driver node and cannot be pickled or transferred to Spark or Ray worker nodes. This error can appear even if your code doesn't directly call dbutils—if any import or dependency (including libraries or initialization scripts) references dbutils at the module/global level, that reference may be serialized along with your training function or objects, causing the error when trainer.fit() is called in a distributed context.

Some tips: 

• Relocate any imports—especially those that might transitively pull in dbutils—inside your train_fn (the function passed to TorchTrainer). This keeps driver-only modules (like dbutils) out of the scope that Ray/TorchTrainer serializes to the workers.
• Make sure neither your code nor any imported script/module at global scope uses dbutils.* anywhere outside the driver before launching distributed jobs.
• Retrieve dbutils-based values on the driver before launching the trainer, then pass them to the workers as standard variables (not as part of module-level imports/objects).

Let me know if this helps! 

Best, 

Sarah