diff --git a/llms/deepseek-coder/convert.py b/llms/deepseek-coder/convert.py
index 06c681c4..f37dbe38 100644
--- a/llms/deepseek-coder/convert.py
+++ b/llms/deepseek-coder/convert.py
@@ -44,7 +44,7 @@ def convert(args):
     config = model.config.to_dict()
 
     state_dict = model.state_dict()
-    tokenizer = AutoTokenizer.from_pretrained(str(hf_path), trust_remote_code=True)
+    tokenizer = AutoTokenizer.from_pretrained(str(hf_path), trust_remote_code=True, use_fast=False)
 
     # things to change
     # 1. there's no "model." in the weight names
@@ -84,7 +84,7 @@ def convert(args):
 
     weights = {k: v.numpy() for k, v in state_dict.items()}
 
-    config["rope_scaling_factor"] = config["rope_scaling"]["factor"]
+    config["rope_scaling_factor"] = config["rope_scaling"]["factor"] if config["rope_scaling"] is not None else 1.0
     keep_keys = set(
         [
             "vocab_size",
@@ -96,6 +96,7 @@ def convert(args):
             "rms_norm_eps",
             "intermediate_size",
             "rope_scaling_factor",
+            "rope_theta"
         ]
     )
     for k in list(config.keys()):
@@ -151,4 +152,4 @@ if __name__ == "__main__":
     tokenizer.save_pretrained(mlx_path)
     with open(mlx_path / "config.json", "w") as f:
         config["model_type"] = "deepseek_coder"
-        json.dump(config, f, indent=4)
+        json.dump(config, f, indent=4)%