use bfloat16 precision for training to halve RAM usage

app.py

@@ -363,7 +363,7 @@ def gradio_main():
                 variant="turbo",
                 max_duration=60.0,
                 device="cpu",
-                precision="float32",
+                precision="bfloat16",
             )
 
             processed = result.get("processed", 0)
@@ -386,9 +386,9 @@ def gradio_main():
                 checkpoint_dir=ACE_CHECKPOINT_DIR,
                 variant="turbo",
                 device="cpu",
-                precision="float32",
+                precision="bfloat16",
             )
-            model = model.float()
+            model = model.bfloat16()
 
             adapter_cfg = LoRAConfigV2(r=rank, alpha=rank, dropout=0.0)
             train_cfg = TrainingConfigV2(
@@ -400,7 +400,7 @@ def gradio_main():
                 batch_size=1,
                 learning_rate=lr,
                 device="cpu",
-                precision="float32",
+                precision="bfloat16",
                 seed=42,
                 num_workers=0,
                 pin_memory=False,