force float32 on training model for CPU

app.py

@@ -304,13 +304,14 @@ def train_lora(
         from acestep.training_v2.trainer_fixed import FixedLoRATrainer
         from acestep.training_v2.configs import TrainingConfigV2, LoRAConfigV2
 
-        # Load model for training
+        # Load model for training (force float32 for CPU)
         model = load_decoder_for_training(
             checkpoint_dir=CHECKPOINT_DIR,
             variant="turbo",
             device="cpu",
             precision="float32",
         )
+        model = model.float()
 
         adapter_cfg = LoRAConfigV2(
             r=rank,