support kimi + more options in chat mode (#1312)

llms/mlx_lm/_version.py

@@ -1,3 +1,3 @@
 # Copyright © 2023-2024 Apple Inc.
 
-__version__ = "0.21.5"
+__version__ = "0.21.6"

llms/mlx_lm/chat.py

@@ -65,12 +65,25 @@ def main():
         tokenizer_config={"trust_remote_code": True},
     )
 
-    print(f"[INFO] Starting chat session with {args.model}. To exit, enter 'q'.")
+    def print_help():
+        print("The command list:")
+        print("- 'q' to exit")
+        print("- 'r' to reset the chat")
+        print("- 'h' to display these commands")
+
+    print(f"[INFO] Starting chat session with {args.model}.")
+    print_help()
     prompt_cache = make_prompt_cache(model, args.max_kv_size)
     while True:
         query = input(">> ")
         if query == "q":
             break
+        if query == "r":
+            prompt_cache = make_prompt_cache(model, args.max_kv_size)
+            continue
+        if query == "h":
+            print_help()
+            continue
         messages = [{"role": "user", "content": query}]
         prompt = tokenizer.apply_chat_template(messages, add_generation_prompt=True)
         for response in stream_generate(

llms/mlx_lm/models/deepseek_v3.py

@@ -181,6 +181,7 @@ class DeepseekV3Attention(nn.Module):
             bias=config.attention_bias,
         )
 
+        if self.config.rope_scaling is not None:
             mscale_all_dim = self.config.rope_scaling.get("mscale_all_dim", 0)
             scaling_factor = self.config.rope_scaling["factor"]
             if mscale_all_dim:
@@ -205,6 +206,12 @@ class DeepseekV3Attention(nn.Module):
                 base=self.rope_theta,
                 **rope_kwargs,
             )
+        else:
+            self.rope = nn.RoPE(
+                dims=self.qk_rope_head_dim,
+                base=self.rope_theta,
+                traditional=True,
+            )
 
     def __call__(
         self,
@@ -487,8 +494,12 @@ class Model(nn.Module):
                         ]
                         weights[f"{prefix}.mlp.switch_mlp.{m}.{k}"] = mx.stack(to_join)
 
-        # Remove multi-token prediction layer
+        # Remove multi-token prediction layer and any unused precomputed rotary freqs
-        return {k: v for k, v in weights.items() if not k.startswith("model.layers.61")}
+        return {
+            k: v
+            for k, v in weights.items()
+            if not k.startswith("model.layers.61") and "rotary_emb.inv_freq" not in k
+        }
 
     @property
     def layers(self):

llms/mlx_lm/utils.py

@@ -191,6 +191,7 @@ def get_model_path(path_or_hf_repo: str, revision: Optional[str] = None) -> Path
                         "*.py",
                         "tokenizer.model",
                         "*.tiktoken",
+                        "tiktoken.model",
                         "*.txt",
                         "*.jsonl",
                     ],