feat: add Llama 3.2 chat template with Cutting Knowledge Date

python/mlc_llm/conversation_template/llama.py

@@ -29,6 +29,38 @@
     )
 )
 
+# Llama3.2 -- same as Llama3.1 but includes "Cutting Knowledge Date" and "Today Date" in system
+# template to match the official HuggingFace chat template.
+# See https://github.com/mlc-ai/mlc-llm/issues/3002
+ConvTemplateRegistry.register_conv_template(
+    Conversation(
+        name="llama-3_2",
+        system_template=(
+            "<|start_header_id|>system<|end_header_id|>\n\n"
+            "Cutting Knowledge Date: December 2023\n"
+            f"Today Date: {MessagePlaceholders.TODAY_DATE.value}\n\n"
+            f"{MessagePlaceholders.SYSTEM.value}<|eot_id|>"
+        ),
+        system_message="You are a helpful, respectful and honest assistant.",
+        roles={
+            "user": "<|start_header_id|>user",
+            "assistant": "<|start_header_id|>assistant",
+            "tool": "<|start_header_id|>ipython",
+        },
+        seps=["<|eot_id|>"],
+        role_content_sep="<|end_header_id|>\n\n",
+        role_empty_sep="<|end_header_id|>\n\n",
+        stop_str=[],
+        stop_token_ids=[
+            128001,
+            128008,
+            128009,
+        ],  # "<|end_of_text|>", "<|eom_id|>", "<|eot_id|>"
+        system_prefix_token_ids=[128000],  # "<|begin_of_text|>"
+        add_role_after_system_message=True,
+    )
+)
+
 # Llama3.1 -- same as Llama3 except stop token ids and stop str
 ConvTemplateRegistry.register_conv_template(
     Conversation(

python/mlc_llm/protocol/conversation_protocol.py

@@ -1,5 +1,6 @@
 """The standard conversation protocol in MLC LLM"""
 
+from datetime import datetime
 from enum import Enum
 from typing import Any, Dict, List, Optional, Tuple, Type, TypeVar, Union
 
@@ -15,6 +16,7 @@ class MessagePlaceholders(Enum):
     ASSISTANT = "{assistant_message}"
     TOOL = "{tool_message}"
     FUNCTION = "{function_string}"
+    TODAY_DATE = "{today_date}"
 
 
 T = TypeVar("T", bound="BaseModel")
@@ -127,6 +129,11 @@ def as_prompt(self, config=None) -> List[Any]:
         system_msg = self.system_template.replace(
             MessagePlaceholders.SYSTEM.value, self.system_message
         )
+        # Replace the today_date placeholder with the current date.
+        system_msg = system_msg.replace(
+            MessagePlaceholders.TODAY_DATE.value,
+            datetime.now().strftime("%d %b %Y"),
+        )
 
         # - Get the message strings.
         message_list: List[Union[str, data.Data]] = []

tests/python/conversation_template/test_llama_template.py

@@ -41,3 +41,50 @@ def test_llama3_prompt():
 
 if __name__ == "__main__":
     test_llama3_prompt()
+
+
+def test_llama3_2_prompt():
+    """Test that Llama 3.2 template includes Cutting Knowledge Date and dynamic Today Date.
+
+    See https://github.com/mlc-ai/mlc-llm/issues/3002
+    """
+    from datetime import datetime
+
+    conversation = ConvTemplateRegistry.get_conv_template("llama-3_2")
+    assert conversation is not None, "llama-3_2 template should be registered"
+
+    system_msg = "You are a helpful assistant."
+    user_msg = "What is the capital of France?"
+
+    conversation.system_message = system_msg
+    conversation.messages.append(("user", user_msg))
+    conversation.messages.append(("assistant", None))
+    res = conversation.as_prompt()
+
+    today = datetime.now().strftime("%d %b %Y")
+    expected = (
+        "<|start_header_id|>system<|end_header_id|>\n\n"
+        "Cutting Knowledge Date: December 2023\n"
+        f"Today Date: {today}\n\n"
+        "You are a helpful assistant.<|eot_id|>"
+        "<|start_header_id|>user<|end_header_id|>\n\n"
+        "What is the capital of France?<|eot_id|>"
+        "<|start_header_id|>assistant<|end_header_id|>\n\n"
+    )
+
+    assert res[0] == expected
+
+
+def test_llama3_2_has_tool_role():
+    """Test that Llama 3.2 template supports tool/ipython role like 3.1."""
+    conversation = ConvTemplateRegistry.get_conv_template("llama-3_2")
+    assert "tool" in conversation.roles
+    assert conversation.roles["tool"] == "<|start_header_id|>ipython"
+
+
+def test_llama3_2_stop_tokens():
+    """Test that Llama 3.2 has the correct stop token IDs."""
+    conversation = ConvTemplateRegistry.get_conv_template("llama-3_2")
+    assert 128001 in conversation.stop_token_ids  # <|end_of_text|>
+    assert 128008 in conversation.stop_token_ids  # <|eom_id|>
+    assert 128009 in conversation.stop_token_ids  # <|eot_id|>