Using a different ref_model from model leads to incorrect results

[DarshanDeshpande]

System Info

• Platform: Linux-6.8.0-47-generic-x86_64-with-glibc2.35
• Python version: 3.10.15
• PyTorch version: 2.4.0
• CUDA device(s): NVIDIA H100 80GB HBM3, NVIDIA H100 80GB HBM3, NVIDIA H100 80GB HBM3, NVIDIA H100 80GB HBM3
• Transformers version: 4.46.1
• Accelerate version: 1.1.0
• Accelerate config:
  • compute_environment: LOCAL_MACHINE
  • distributed_type: FSDP
  • mixed_precision: bf16
  • use_cpu: False
  • debug: True
  • num_processes: 4
  • machine_rank: 0
  • num_machines: 1
  • rdzv_backend: static
  • same_network: True
  • main_training_function: main
  • enable_cpu_affinity: False
  • fsdp_config: {'fsdp_activation_checkpointing': True, 'fsdp_auto_wrap_policy': 'TRANSFORMER_BASED_WRAP', 'fsdp_backward_prefetch': 'BACKWARD_PRE', 'fsdp_cpu_ram_efficient_loading': True, 'fsdp_forward_prefetch': True, 'fsdp_offload_params': False, 'fsdp_sharding_strategy': 'FULL_SHARD', 'fsdp_state_dict_type': 'FULL_STATE_DICT', 'fsdp_sync_module_states': True, 'fsdp_use_orig_params': True}
  • downcast_bf16: no
  • tpu_use_cluster: False
  • tpu_use_sudo: False
  • tpu_env: []
• Datasets version: 3.1.0
• HF Hub version: 0.26.2
• TRL version: 0.12.0
• 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: not installed
• PEFT version: not installed

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

I noticed that the DPO trainer uses the processing_class to tokenize inputs to both model and ref_model. Is there a way to allow for a different ref_model base class that does not share the same tokenizer config with the model? For example using a Llama-3.1-8b model to align a Llama-3.2-3b model - training with this configuration leads to a constant loss=1.0 at the moment.

Expected behavior

The Trainer must take two processing classes and allow for a different ref_model and model class

[qgallouedec]

Indeed, it's not currently supported. And unless it's widely demanded, I don't think it will be.

Having said that, I think you can easily implement it. The following should work:

1. set precompute_ref_log_probs=True in DPOConfig

2. add a new parameter ref_processing_class in DPOTrainer

3. in DPOTrainer.__init__, create a new tokenized dataset with ref_processing_class something like

   fn_kwargs = {
       "processing_class": ref_processing_class,  # <-
       "max_prompt_length": args.max_prompt_length,
       "max_completion_length": args.max_completion_length,
       # for enc-dec, we add the special tokens ([bos_token] + prompt + [eos_token]; completion + [eos_token])
       "add_special_tokens": self.is_encoder_decoder,
   }
   self.ref_train_dataset = train_dataset.map(  # <-
       self.tokenize_row if not self.is_vision_model else self.process_row,
       fn_kwargs=fn_kwargs,
       num_proc=self.dataset_num_proc,
       writer_batch_size=10,
       desc="Tokenizing train dataset",
   )

4. modify precomputing ref part here

   trl/trl/trainer/dpo_trainer.py

   Lines 682 to 692 in 74e20cb

   	if self.precompute_ref_log_probs and not self._precomputed_train_ref_log_probs:
   	dataloader_params = {
   	"batch_size": self.args.per_device_train_batch_size,
   	"collate_fn": self.data_collator,
   	"num_workers": self.args.dataloader_num_workers,
   	"pin_memory": self.args.dataloader_pin_memory,
   	"shuffle": False,
   	}
   	# prepare dataloader
   	data_loader = self.accelerator.prepare(DataLoader(self.train_dataset, **dataloader_params))

   with

   data_loader = self.accelerator.prepare(DataLoader(self.ref_train_dataset, **dataloader_params))