bump trl and accelerate for latest releases (#1730)

* bump trl and accelerate for latest releases

* ensure that the CI runs on new gh org

* drop kto_pair support since removed upstream

src/axolotl/core/trainer_builder.py

@@ -1670,8 +1670,6 @@ class HFRLTrainerBuilder(TrainerBuilderBase):
             dpo_trainer_kwargs["loss_type"] = "ipo"
             if self.cfg.dpo_label_smoothing:
                 dpo_trainer_kwargs["label_smoothing"] = self.cfg.dpo_label_smoothing
-        elif self.cfg.rl == "kto_pair":
-            dpo_trainer_kwargs["loss_type"] = "kto_pair"
         if self.eval_dataset:
             dpo_trainer_kwargs["eval_dataset"] = self.eval_dataset
         if self.cfg.adapter and self.peft_config:
@@ -1680,7 +1678,7 @@ class HFRLTrainerBuilder(TrainerBuilderBase):
             dpo_trainer_kwargs[
                 "precompute_ref_log_probs"
             ] = self.cfg.precompute_ref_log_probs
-        if self.cfg.rl in ["dpo", "ipo", "kto_pair"]:
+        if self.cfg.rl in ["dpo", "ipo"]:
             trainer_cls = AxolotlDPOTrainer
             dpo_trainer_kwargs["beta"] = self.cfg.rl_beta or 0.1
             trainer_cls_args = [self.model, self.model_ref]
@@ -1695,7 +1693,7 @@ class HFRLTrainerBuilder(TrainerBuilderBase):
         elif self.cfg.rl == "orpo":
             trainer_cls = AxolotlORPOTrainer
             trainer_cls_args = [self.model]
-        elif self.cfg.rl == "kto":
+        elif self.cfg.rl in ["kto"]:
             trainer_cls = AxolotlKTOTrainer
             trainer_cls_args = [self.model]
         else:

src/axolotl/utils/config/models/input/v0_4_1/__init__.py

@@ -165,7 +165,6 @@ class RLType(str, Enum):
 
     dpo = "dpo"  # pylint: disable=invalid-name
     ipo = "ipo"  # pylint: disable=invalid-name
-    kto_pair = "kto_pair"  # pylint: disable=invalid-name
     orpo = "orpo"  # pylint: disable=invalid-name
     kto = "kto"  # pylint: disable=invalid-name
 

src/axolotl/utils/models.py

@@ -805,11 +805,7 @@ def load_model(
     if not reference_model or cfg.lora_model_dir:
         # if we're not loading the reference model, then we're loading the model for training
         # then the dpo trainer doesn't want the peft model loaded over it, it just wants the lora/peft config
-        if (
-            cfg.adapter
-            and cfg.rl in ["dpo", "ipo", "kto_pair", "kto"]
-            and not cfg.merge_lora
-        ):
+        if cfg.adapter and cfg.rl in ["dpo", "ipo", "kto"] and not cfg.merge_lora:
             _, lora_config = load_lora(model, cfg, inference=False, config_only=True)
         else:
             model, lora_config = load_adapter(model, cfg, cfg.adapter)

src/axolotl/utils/trainer.py

@@ -427,7 +427,7 @@ def prepare_optim_env(cfg):
 
 
 def setup_trainer(cfg, train_dataset, eval_dataset, model, tokenizer, total_num_steps):
-    if cfg.rl in ["dpo", "ipo", "kto_pair", "orpo", "kto"]:
+    if cfg.rl in ["dpo", "ipo", "orpo", "kto"]:
         trainer_builder = HFRLTrainerBuilder(cfg, model[0], tokenizer)
         trainer_builder.model_ref = model[1]
         trainer_builder.peft_config = model[2]