feat: add handling for seed and SP/ring-attn config

src/axolotl/core/trainer_builder.py

@@ -331,6 +331,9 @@ class TrainerBuilderBase(abc.ABC):
             "save_only_model",
             "include_tokens_per_second",
             "weight_decay",
+            "sequence_parallel_degree",
+            "ring_attn_func",
+            "seed",
         ]:
             if hasattr(self.cfg, arg) and getattr(self.cfg, arg) is not None:
                 training_args_kwargs[arg] = getattr(self.cfg, arg)
@@ -593,9 +596,6 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
     def build(self, total_num_steps):
         training_arguments_kwargs = self._set_base_training_args(total_num_steps)
 
-        if self.cfg.seed is not None:
-            training_arguments_kwargs["seed"] = self.cfg.seed
-
         if self.cfg.fsdp:
             training_arguments_kwargs["fsdp"] = self.cfg.fsdp
             if self.cfg.fsdp_config:
@@ -806,11 +806,6 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
                 self.cfg.kd_top_k_before_softmax
             )
 
-        training_arguments_kwargs["sequence_parallel_degree"] = (
-            self.cfg.sequence_parallel_degree
-        )
-        training_arguments_kwargs["ring_attn_func"] = self.cfg.ring_attn_func
-
         if self.cfg.reward_model:
             training_args_cls = AxolotlRewardConfig
         elif self.cfg.process_reward_model:
@@ -1010,10 +1005,6 @@ class HFRLTrainerBuilder(TrainerBuilderBase):
         if self.cfg.use_wandb:
             training_args_kwargs["run_name"] = self.cfg.wandb_name
 
-        training_args_kwargs["sequence_parallel_degree"] = (
-            self.cfg.sequence_parallel_degree
-        )
-
         training_args_cls = None
         blocklist_args_kwargs = []
         if self.cfg.rl is RLType.SIMPO: