diff --git a/examples/llama.py b/examples/llama.py
index fefe248ed5..4139a91e88 100755
--- a/examples/llama.py
+++ b/examples/llama.py
@@ -262,8 +262,8 @@ class AbsmaxQuantizedLinear:
       if 'feed_forward' in name or ('attention.w') in name or name == 'output.weight':
         scale = v.abs().max(axis=1) / 127.0
         int8_weight = (v.T/scale).T.cast(dtype=dtypes.int8)
-        new_tensors[name] = int8_weight.realize()
-        new_tensors[name.replace('weight', 'scale')] = scale.realize()
+        new_tensors[name] = int8_weight
+        new_tensors[name.replace('weight', 'scale')] = scale
       else:
         new_tensors[name] = v
     return new_tensors
@@ -287,6 +287,7 @@ class LLaMa:
 
     if quantize:
       weights = AbsmaxQuantizedLinear.quantize(weights)
+      for _,v in weights.items(): v.realize()
     load_state_dict(model, weights, strict=False)
 
     return LLaMa(model, sp_model)