Fix fusion and test

Author: awni

mlx_lm/fuse.py

@@ -78,6 +78,7 @@ def main() -> None:
         print("Dequantizing model")
         model = dequantize_model(model)
         config.pop("quantization", None)
+        config.pop("quantization_config", None)
 
     save_path = Path(args.save_path)
     save(

mlx_lm/models/deepseek_v32.py

@@ -224,7 +224,7 @@ def __call__(
             mask = sparse_mask
         # Ensure the indexer cache is evaluated even if the topk_indices are unused
         # to keep the graph from getting too large
-        if cache is not None:
+        if cache is not None and cache[0] is not None:
             cache[0].keys = mx.depends(cache[0].keys, (cache[1].keys, cache[1].values))
 
         output = scaled_dot_product_attention(