[Fix] Fix deepseek r1 stop tokens (#3098)

This PR removes `deepseek_v3` conversation template, and adds
`deepseek_r1_qwen` and `deepseek_r1_llama`, which only differ
by the stop token. Note that DeepSeek V3 has its own stop token id,
different from the R1's.

Author: CharlieFRuan

python/mlc_llm/conversation_template/deepseek.py

@@ -36,10 +36,10 @@
     )
 )
 
-# Deepseek V3
+# DeepSeek-R1-Distill-Qwen
 ConvTemplateRegistry.register_conv_template(
     Conversation(
-        name="deepseek_v3",
+        name="deepseek_r1_qwen",
         system_template=f"<｜begin▁of▁sentence｜>{MessagePlaceholders.SYSTEM.value}",
         system_message="You are a helpful assistant.",
         roles={"user": "<｜User｜>", "assistant": "<｜Assistant｜>"},
@@ -49,3 +49,17 @@
         stop_token_ids=[151643],
     )
 )
+
+# DeepSeek-R1-Distill-Llama, exactly the same as DeepSeek-R1-Distill-Qwen, but different stop token
+ConvTemplateRegistry.register_conv_template(
+    Conversation(
+        name="deepseek_r1_llama",
+        system_template=f"<｜begin▁of▁sentence｜>{MessagePlaceholders.SYSTEM.value}",
+        system_message="You are a helpful assistant.",
+        roles={"user": "<｜User｜>", "assistant": "<｜Assistant｜>"},
+        seps=["", "<｜end▁of▁sentence｜>"],
+        role_content_sep="",
+        role_empty_sep="",
+        stop_token_ids=[128001],
+    )
+)

python/mlc_llm/interface/gen_config.py

@@ -309,7 +309,8 @@ def gen_config(  # pylint: disable=too-many-locals,too-many-arguments,too-many-b
     "aya-23",
     "deepseek",
     "deepseek_v2",
-    "deepseek_v3",
+    "deepseek_r1_qwen",
+    "deepseek_r1_llama",
     "olmo",
     "nemotron",
 }