diff --git a/bark/generation.py b/bark/generation.py
index ec313e7..4ac165c 100644
--- a/bark/generation.py
+++ b/bark/generation.py
@@ -298,6 +298,7 @@ def load_model(use_gpu=True, use_small=False, force_reload=False, model_type="te
     if model_type not in ("text", "coarse", "fine"):
         raise NotImplementedError()
     global models
+    global models_devices
     device = _grab_best_device(use_gpu=use_gpu)
     model_key = f"{model_type}"
     if OFFLOAD_CPU:
@@ -317,6 +318,7 @@ def load_model(use_gpu=True, use_small=False, force_reload=False, model_type="te
 
 def load_codec_model(use_gpu=True, force_reload=False):
     global models
+    global models_devices
     device = _grab_best_device(use_gpu=use_gpu)
     if device == "mps":
         # encodec doesn't support mps
@@ -421,6 +423,7 @@ def generate_text_semantic(
         semantic_history = None
     # load models if not yet exist
     global models
+    global models_devices
     if "text" not in models:
         preload_models()
     model_container = models["text"]
@@ -616,6 +619,7 @@ def generate_coarse(
         x_coarse_history = np.array([], dtype=np.int32)
     # load models if not yet exist
     global models
+    global models_devices
     if "coarse" not in models:
         preload_models()
     model = models["coarse"]
@@ -755,6 +759,7 @@ def generate_fine(
     n_coarse = x_coarse_gen.shape[0]
     # load models if not yet exist
     global models
+    global models_devices
     if "fine" not in models:
         preload_models()
     model = models["fine"]
@@ -842,6 +847,7 @@ def codec_decode(fine_tokens):
     """Turn quantized audio codes into audio array using encodec."""
     # load models if not yet exist
     global models
+    global models_devices
     if "codec" not in models:
         preload_models()
     model = models["codec"]