Support jinja as chat template file (#1104)

python/sglang/srt/openai_api/adapter.py

@@ -117,7 +117,7 @@ def create_streaming_error_response(
     return json_str
 
 
-def load_chat_template_for_openai_api(chat_template_arg):
+def load_chat_template_for_openai_api(tokenizer_manager, chat_template_arg):
     global chat_template_name
 
     print(f"Use chat template: {chat_template_arg}")
@@ -127,6 +127,17 @@ def load_chat_template_for_openai_api(chat_template_arg):
                 f"Chat template {chat_template_arg} is not a built-in template name "
                 "or a valid chat template file path."
             )
+        if chat_template_arg.endswith(".jinja"):
+            with open(chat_template_arg, "r") as f:
+                chat_template = "".join(f.readlines()).strip("\n")
+            tokenizer_manager.tokenizer.chat_template = chat_template.replace(
+                "\\n", "\n"
+            )
+            chat_template_name = None
+        else:
+            assert chat_template_arg.endswith(
+                ".json"
+            ), "unrecognized format of chat template file"
             with open(chat_template_arg, "r") as filep:
                 template = json.load(filep)
                 try:

python/sglang/srt/server.py

@@ -288,6 +288,8 @@ def launch_server(
 
     # Launch processes
     tokenizer_manager = TokenizerManager(server_args, port_args, model_overide_args)
+    if server_args.chat_template:
+        load_chat_template_for_openai_api(tokenizer_manager, server_args.chat_template)
     pipe_controller_reader, pipe_controller_writer = mp.Pipe(duplex=False)
     pipe_detoken_reader, pipe_detoken_writer = mp.Pipe(duplex=False)
 
@@ -375,11 +377,6 @@ def _set_envs_and_config(server_args: ServerArgs):
         # FIXME: remove this after https://github.com/triton-lang/triton/pull/4295 is used as a dependency.
         maybe_set_triton_cache_manager()
 
-    # Set global chat template
-    if server_args.chat_template:
-        # TODO: replace this with huggingface transformers template
-        load_chat_template_for_openai_api(server_args.chat_template)
-
     # Check flashinfer version
     if not server_args.disable_flashinfer:
         assert_pkg_version(