--auto_batch_finding is silently ignored for torchFSDP trainer

[Demirrr]

Summary

--auto_batch_finding is documented (via its --help text) as "Only available for TP trainer", but in practice it is also wired up for the single-GPU/CPU trainer and the DDP trainer — just not for FSDP. When combined with --trainer torchFSDP, the flag is silently accepted and has no effect: TorchFSDPTrainer never reads self.attributes.auto_batch_finding and just uses --batch_size as given.

Details

• Flag definition: dicee/config.py:171 (self.auto_batch_finding: bool = False) and dicee/scripts/run.py:153 (--auto_batch_finding, help: "Find a batch size fitting in GPUs. Only available for TP trainer").
• Implemented in:
  • dicee/trainer/torch_trainer.py:92-109 (single-GPU/CPU)
  • dicee/trainer/torch_trainer_ddp.py:134-149 (DDP)
  • dicee/trainer/model_parallelism.py:77-92 (TP)
• Not implemented in dicee/trainer/torch_trainer_fsdp.py — no reference to auto_batch_finding anywhere in the file. It reads self.attributes.batch_size directly (line 202) and never attempts auto-sizing.

Impact

Running dicee --trainer torchFSDP --auto_batch_finding ... gives no error or warning; users may assume batch size is being auto-tuned when it is not, and must set --batch_size manually for FSDP runs.

Suggested fix

Either:

1. Wire auto_batch_finding into TorchFSDPTrainer (analogous to the DDP implementation, respecting rank-0 discovery + broadcast), or
2. Explicitly warn/raise when --auto_batch_finding is combined with --trainer torchFSDP, and update the --help text/README to reflect actual trainer support instead of only mentioning TP.