fix: DeepSpeed device_placement ValueError

[hjinnkim]

accelerator.prepare() fails with DeepSpeed due to custom device_placement parameter.

Problem

When using DeepSpeed ZeRO Stage 2/3, training fails with:

ValueError: You can't customize device placements with DeepSpeed or Megatron-LM.

Root Cause

Accelerate raises this error when device_placement is passed with DeepSpeed:

• accelerate/accelerator.py:1458-1461

SimpleTuner passes device_placement unconditionally:

• trainer.py:3628-3632

Solution

Skip device_placement parameter when using DeepSpeed, as DeepSpeed handles device placement internally.

Test

• DeepSpeed ZeRO Stage 2, 8 GPUs - training starts without error