↩️ Revert tokenizer hotfix #2163

trl/trainer/bco_trainer.py

@@ -699,16 +699,6 @@ def make_inputs_require_grad(module, input, output):
             embeddings.cpu().float().numpy(), labels.cpu().numpy()
         )
 
-    # Hot fix to avoid error when setting tokenizer after https://github.com/huggingface/transformers/pull/32385
-    # Should be removed when fixed in transformers, or whenhttps://github.com/huggingface/trl/pull/2162 is merged.
-    @property
-    def tokenizer(self):
-        return self.processing_class
-
-    @tokenizer.setter
-    def tokenizer(self, tokenizer):
-        self.processing_class = tokenizer
-
     @property
     def match_underlying_distribution(self):
         return self.embedding_func is not None and self.embedding_tokenizer is not None

trl/trainer/cpo_trainer.py

@@ -330,16 +330,6 @@ def make_inputs_require_grad(module, input, output):
                 "Your `Trainer` does not have an `accelerator` object. Consider upgrading `transformers`."
             )
 
-    # Hot fix to avoid error when setting tokenizer after https://github.com/huggingface/transformers/pull/32385
-    # Should be removed when fixed in transformers, or whenhttps://github.com/huggingface/trl/pull/2162 is merged.
-    @property
-    def tokenizer(self):
-        return self.processing_class
-
-    @tokenizer.setter
-    def tokenizer(self, tokenizer):
-        self.processing_class = tokenizer
-
     def build_tokenized_answer(self, prompt, answer):
         """
         Llama tokenizer does satisfy `enc(a + b) = enc(a) + enc(b)`.

trl/trainer/dpo_trainer.py

@@ -907,16 +907,6 @@ def make_inputs_require_grad(module, input, output):
         if self.loss_type == "bco_pair":
             self.running = RunningMoments(self.accelerator)
 
-    # Hot fix to avoid error when setting tokenizer after https://github.com/huggingface/transformers/pull/32385
-    # Should be removed when fixed in transformers, or whenhttps://github.com/huggingface/trl/pull/2162 is merged.
-    @property
-    def tokenizer(self):
-        return self.processing_class
-
-    @tokenizer.setter
-    def tokenizer(self, tokenizer):
-        self.processing_class = tokenizer
-
     def _prepare_deepspeed(self, model: PreTrainedModelWrapper):
         # Adapted from accelerate: https://github.com/huggingface/accelerate/blob/739b135f8367becb67ffaada12fe76e3aa60fefd/src/accelerate/accelerator.py#L1473
         deepspeed_plugin = self.accelerator.state.deepspeed_plugin

trl/trainer/iterative_sft_trainer.py

@@ -163,16 +163,6 @@ def __init__(
 
         PPODecorators.optimize_device_cache = self.optimize_device_cache
 
-    # Hot fix to avoid error when setting tokenizer after https://github.com/huggingface/transformers/pull/32385
-    # Should be removed when fixed in transformers, or whenhttps://github.com/huggingface/trl/pull/2162 is merged.
-    @property
-    def tokenizer(self):
-        return self.processing_class
-
-    @tokenizer.setter
-    def tokenizer(self, tokenizer):
-        self.processing_class = tokenizer
-
     def prepare_model_inputs(self, input_ids: torch.Tensor, attention_mask: torch.Tensor, labels: torch.Tensor):
         if attention_mask is None:
             attention_mask = [torch.ones_like(ids) for ids in input_ids]

trl/trainer/kto_trainer.py

@@ -720,16 +720,6 @@ def make_inputs_require_grad(module, input, output):
             else:
                 self.ref_model = self.accelerator.prepare_model(self.ref_model, evaluation_mode=True)
 
-    # Hot fix to avoid error when setting tokenizer after https://github.com/huggingface/transformers/pull/32385
-    # Should be removed when fixed in transformers, or whenhttps://github.com/huggingface/trl/pull/2162 is merged.
-    @property
-    def tokenizer(self):
-        return self.processing_class
-
-    @tokenizer.setter
-    def tokenizer(self, tokenizer):
-        self.processing_class = tokenizer
-
     def _prepare_deepspeed(self, model: PreTrainedModelWrapper):
         # Adapted from accelerate: https://github.com/huggingface/accelerate/blob/739b135f8367becb67ffaada12fe76e3aa60fefd/src/accelerate/accelerator.py#L1473
         deepspeed_plugin = self.accelerator.state.deepspeed_plugin

trl/trainer/orpo_trainer.py

@@ -351,16 +351,6 @@ def _prepare_deepspeed(self, model: PreTrainedModelWrapper):
         model.eval()
         return model
 
-    # Hot fix to avoid error when setting tokenizer after https://github.com/huggingface/transformers/pull/32385
-    # Should be removed when fixed in transformers, or whenhttps://github.com/huggingface/trl/pull/2162 is merged.
-    @property
-    def tokenizer(self):
-        return self.processing_class
-
-    @tokenizer.setter
-    def tokenizer(self, tokenizer):
-        self.processing_class = tokenizer
-
     def build_tokenized_answer(self, prompt, answer):
         """
         Llama tokenizer does satisfy `enc(a + b) = enc(a) + enc(b)`.

trl/trainer/ppov2_trainer.py

@@ -243,16 +243,6 @@ def __init__(
             self.ref_policy = self.ref_policy.to(self.accelerator.device)
             self.reward_model = self.reward_model.to(self.accelerator.device)
 
-    # Hot fix to avoid error when setting tokenizer after https://github.com/huggingface/transformers/pull/32385
-    # Should be removed when fixed in transformers, or whenhttps://github.com/huggingface/trl/pull/2162 is merged.
-    @property
-    def tokenizer(self):
-        return self.processing_class
-
-    @tokenizer.setter
-    def tokenizer(self, tokenizer):
-        self.processing_class = tokenizer
-
     def get_train_dataloader(self) -> DataLoader:
         return self.dataloader
 

trl/trainer/rloo_trainer.py

@@ -216,16 +216,6 @@ def __init__(
             self.ref_policy = self.ref_policy.to(self.accelerator.device)
             self.reward_model = self.reward_model.to(self.accelerator.device)
 
-    # Hot fix to avoid error when setting tokenizer after https://github.com/huggingface/transformers/pull/32385
-    # Should be removed when fixed in transformers, or whenhttps://github.com/huggingface/trl/pull/2162 is merged.
-    @property
-    def tokenizer(self):
-        return self.processing_class
-
-    @tokenizer.setter
-    def tokenizer(self, tokenizer):
-        self.processing_class = tokenizer
-
     def get_train_dataloader(self) -> DataLoader:
         return self.dataloader