System Info

• Platform: Linux-5.15.0-1061-gke-x86_64-with-glibc2.31
• Python version: 3.11.9
• PyTorch version: 2.4.0
• CUDA device(s): NVIDIA A100-SXM4-80GB
• Transformers version: 4.46.3
• Accelerate version: 1.0.1
• Accelerate config: not found
• Datasets version: 3.0.2
• HF Hub version: 0.27.0
• TRL version: 0.12.1
• bitsandbytes version: 0.44.1
• DeepSpeed version: not installed
• Diffusers version: not installed
• Liger-Kernel version: not installed
• LLM-Blender version: not installed
• OpenAI version: 1.58.1
• PEFT version: 0.13.2

Information

• The official example scripts
• My own modified scripts

Tasks

• An officially supported task in the examples folder
• My own task or dataset (give details below)

Reproduction

difference in how each trainer handles PEFT + FSDP

sft:
https://github.com/huggingface/trl/blob/v0.12.1/trl/trainer/sft_trainer.py#L242-L244

dpo:
https://github.com/huggingface/trl/blob/v0.12.1/trl/trainer/dpo_trainer.py#L363

Expected behavior

Currently workflow is:

• create PEFT model outside of trainer
• pass PEFT model to trainer
• first run SFTTrainer
• use output model from SFTTrainer as base model in DPOTrainer

it is unclear what is the expected way to create and pass a PEFT model to the trainer when also using FSDP for model parallel training since both SFTTrainer and DPOTrainer handle this differently.

Checklist

• I have checked that my issue isn't already filed (see open issues)
• I have included my system information
• Any code provided is minimal, complete, and reproducible (more on MREs)
• Any code provided is properly formatted in code blocks, (no screenshot, more on code blocks)
• Any traceback provided is complete