🤫 TR-DPO implementation

docs/source/dpo_trainer.mdx

@@ -115,6 +115,8 @@ The [SPPO](https://arxiv.org/abs/2405.00675) authors claim that SPPO is capable
 
 The [NCA](https://arxiv.org/abs/2402.05369) authors shows that NCA optimizes the absolute likelihood for each response rather than the relative likelihood.
 
+The [TR-DPO](https://arxiv.org/pdf/2404.09656) paper suggests syncing the reference model weights after every `ref_model_sync_steps` steps of SGD with weight `ref_model_mixup_alpha` during DPO training. To toggle this callback use the `sync_ref_model` flag in the `DPOConfig`.
+
 ## Logging
 
 While training and evaluating we record the following reward metrics:

tests/test_dpo_trainer.py

@@ -308,6 +308,48 @@ def test_dpo_trainer_w_dataset_num_proc(self):
 
                 trainer.train()
 
+    def test_tr_dpo_trainer(self):
+        with tempfile.TemporaryDirectory() as tmp_dir:
+            training_args = DPOConfig(
+                output_dir=tmp_dir,
+                per_device_train_batch_size=2,
+                max_steps=3,
+                remove_unused_columns=False,
+                gradient_accumulation_steps=4,
+                learning_rate=9e-1,
+                evaluation_strategy="steps",
+                precompute_ref_log_probs=False,
+                sync_ref_model=True,
+                ref_model_mixup_alpha=0.5,
+                ref_model_sync_steps=1,
+            )
+
+            dummy_dataset = self._init_dummy_dataset()
+
+            trainer = DPOTrainer(
+                model=self.model,
+                ref_model=self.model,
+                beta=0.1,
+                args=training_args,
+                tokenizer=self.tokenizer,
+                train_dataset=dummy_dataset,
+                eval_dataset=dummy_dataset,
+            )
+
+            # params of the ref model as its the same as the model
+            previous_trainable_params = {n: param.clone() for n, param in trainer.model.named_parameters()}
+
+            trainer.train()
+
+            assert trainer.state.log_history[-1]["train_loss"] is not None
+
+            # check the params have changed
+            for n, param in previous_trainable_params.items():
+                new_param = trainer.ref_model.get_parameter(n)
+                # check the ref model's params have changed - ignore 0 biases
+                if param.sum() != 0:
+                    assert not torch.equal(param, new_param)
+
     @require_no_wandb
     def test_dpo_trainer_generate_during_eval_no_wandb(self):
         with tempfile.TemporaryDirectory() as tmp_dir:

trl/trainer/dpo_config.py

@@ -65,6 +65,12 @@ class DPOConfig(TrainingArguments):
             If True, we ignore the _provided_ reference model and implicitly use a reference model that assigns equal probability to all responses.
         force_use_ref_model (`bool`, defaults to `False`):
             In case one passes a PEFT model for the active model and you want to use a different model for the ref_model, set this flag to `True`.
+        sync_ref_model ('bool', defaults to `False`):
+            The flag for syncing reference model during training from the [TR-DPO](https://arxiv.org/pdf/2404.09656) paper.
+        ref_model_mixup_alpha ('float', defaults to 1.0):
+            The alpha parameter from the [TR-DPO](https://arxiv.org/pdf/2404.09656) paper.
+        ref_model_sync_steps ('int', defaults to 2):
+            The tau parameter from the [TR-DPO](https://arxiv.org/pdf/2404.09656) paper.
     """
 
     beta: float = 0.1
@@ -87,3 +93,6 @@ class DPOConfig(TrainingArguments):
     ref_adapter_name: Optional[str] = None
     reference_free: bool = False
     force_use_ref_model: bool = False
+    sync_ref_model: bool = False
+    ref_model_mixup_alpha: float = 0.9
+    ref_model_sync_steps: int = 64

trl/trainer/dpo_trainer.py

@@ -46,6 +46,7 @@
 from .utils import (
     DPODataCollatorWithPadding,
     RunningMoments,
+    SyncRefModelCallback,
     disable_dropout_in_model,
     pad_to_length,
     peft_module_casting_to_bf16,
@@ -528,12 +529,23 @@ def make_inputs_require_grad(module, input, output):
                 raise ValueError(
                     "No reference model and model is not a Peft model. Try setting `precompute_ref_log_probs=True`"
                 )
+            if args.sync_ref_model:
+                raise ValueError(
+                    "You currently cannot use `ref_model=None` with TR-DPO method. Please provide `ref_model`."
+                )
         else:
             if self.is_deepspeed_enabled:
                 self.ref_model = self._prepare_deepspeed(self.ref_model)
             else:
                 self.ref_model = self.accelerator.prepare_model(self.ref_model, evaluation_mode=True)
 
+        if args.sync_ref_model:
+            if precompute_ref_log_probs:
+                raise ValueError(
+                    "You cannot use `precompute_ref_log_probs=True` with TR-DPO method. Please set `precompute_ref_log_probs=False`."
+                )
+
+            self.add_callback(SyncRefModelCallback(ref_model=self.ref_model, accelerator=self.accelerator))
         if self.loss_type == "bco_pair":
             self.running = RunningMoments(self.accelerator)
 

trl/trainer/utils.py

@@ -20,8 +20,9 @@
 import numpy as np
 import pandas as pd
 import torch
-from accelerate import PartialState
-from accelerate.state import AcceleratorState
+from accelerate import Accelerator
+from accelerate.state import AcceleratorState, PartialState
+from accelerate.utils import is_deepspeed_available
 from rich.console import Console, Group
 from rich.live import Live
 from rich.panel import Panel
@@ -32,6 +33,7 @@
 from transformers import (
     BitsAndBytesConfig,
     DataCollatorForLanguageModeling,
+    PreTrainedModel,
     PreTrainedTokenizerBase,
 )
 from transformers.trainer import TrainerCallback
@@ -45,6 +47,10 @@
     from peft import LoraConfig, PeftConfig
 
 
+if is_deepspeed_available():
+    import deepspeed
+
+
 class AdaptiveKLController:
     """
     Adaptive KL controller described in the paper:
@@ -63,6 +69,39 @@ def update(self, current, n_steps):
         self.value *= mult
 
 
+class SyncRefModelCallback(TrainerCallback):
+    def __init__(
+        self,
+        ref_model: Union[PreTrainedModel, torch.nn.Module],
+        accelerator: Optional[Accelerator],
+    ):
+        self.accelerator = accelerator
+        self.ref_model = ref_model
+
+    @staticmethod
+    def _sync_target_model(model, target_model, alpha):
+        for target_param, copy_param in zip(target_model.parameters(), model.parameters()):
+            target_param.data.mul_(1.0 - alpha).add_(copy_param.data, alpha=alpha)
+
+    @staticmethod
+    def sync_target_model(model, target_model, alpha):
+        deepspeed_plugin = AcceleratorState().deepspeed_plugin
+        if deepspeed_plugin is not None and deepspeed_plugin.zero_stage == 3:
+            with deepspeed.zero.GatheredParameters(list(model.parameters()), modifier_rank=0):
+                if deepspeed.comm.get_rank() == 0:
+                    SyncRefModelCallback._sync_target_model(model, target_model, alpha)
+        else:
+            SyncRefModelCallback._sync_target_model(model, target_model, alpha)
+
+    def on_step_end(self, args, state, control, **kwargs):
+        model: PreTrainedModel = kwargs["model"]
+
+        if self.ref_model is not None and state.global_step % args.ref_model_sync_steps == 0:
+            if self.accelerator:
+                model = self.accelerator.unwrap_model(model)
+            self.sync_target_model(model, self.ref_model, args.ref_model_mixup_alpha)
+
+
 class FixedKLController:
     """Fixed KL controller."""