diff --git a/basaran/model.py b/basaran/model.py
index bfc7be0a..b681cf30 100644
--- a/basaran/model.py
+++ b/basaran/model.py
@@ -323,8 +323,8 @@ def load_model(
         kwargs["device_map"] = "auto"
         kwargs["load_in_8bit"] = load_in_8bit
 
-        # Override the dtype to float16 as required by bitsandbytes.
-        if load_in_8bit:
+        # Cast all parameters to float16 if quantization is enabled.
+        if half_precision or load_in_8bit:
             kwargs["torch_dtype"] = torch.float16
 
     # Support both decoder-only and encoder-decoder models.
@@ -333,10 +333,6 @@ def load_model(
     except ValueError:
         model = AutoModelForSeq2SeqLM.from_pretrained(name_or_path, **kwargs)
 
-    # Cast all parameters to half-precision if required.
-    if half_precision:
-        model = model.half()
-
     # Check if the model has text generation capabilities.
     if not model.can_generate():
         raise TypeError(f"{name_or_path} is not a text generation model")