feat(server): helpers endpoints for conversation format (#613)

* feat: add support for helpers conversation conversion endpoint

also correct schema generation for openllm client

Signed-off-by: Aaron <29749331+aarnphm@users.noreply.github.com>

* chore: update clients to reuse `openllm-core` logics

Signed-off-by: Aaron <29749331+aarnphm@users.noreply.github.com>

* chore: add changelog

Signed-off-by: Aaron <29749331+aarnphm@users.noreply.github.com>

---------

Signed-off-by: Aaron <29749331+aarnphm@users.noreply.github.com>

openllm-python/src/openllm/_service.py

@@ -67,5 +67,35 @@ def metadata_v1(_: str) -> openllm.MetadataOutput:
   return _Metadata
 
 
+class MessagesConverterInput(t.TypedDict):
+  add_generation_prompt: bool
+  messages: t.List[t.Dict[str, t.Any]]
+
+
+class MessageParam(t.TypedDict):
+  role: t.Literal['system', 'user', 'assistant']
+  content: str
+
+
+@svc.api(
+  route='/v1/helpers/messages',
+  input=JSON.from_sample(
+    MessagesConverterInput(
+      add_generation_prompt=False,
+      messages=[
+        MessageParam(role='system', content='You are acting as Ernest Hemmingway.'),
+        MessageParam(role='user', content='Hi there!'),
+        MessageParam(role='assistant', content='Yes?'),
+      ],
+    )
+  ),
+  output=Text(),
+)
+def helpers_messages_v1(message: MessagesConverterInput) -> str:
+  add_generation_prompt = message['add_generation_prompt']
+  messages = message['messages']
+  return llm.tokenizer.apply_chat_template(messages, add_generation_prompt=add_generation_prompt, tokenize=False)
+
+
 # HACK: This must always be the last line in this file, as we will do some MK for OpenAPI schema.
 openllm.mount_entrypoints(svc, llm)

openllm-python/src/openllm/utils/__init__.py

@@ -93,7 +93,7 @@ __all__ = ['generate_labels', 'available_devices', 'device_count']
 
 
 def __dir__() -> t.Sequence[str]:
-  return sorted(__all__)
+  return sorted(__all__) + sorted(dir(openllm_core.utils))
 
 
 def __getattr__(it: str) -> t.Any: