doc: emphasize and notify the usage of chat_template (#3589)

Co-authored-by: Chayenne <zhaochen20@outlook.com>

docs/backend/openai_api_vision.ipynb

@@ -24,7 +24,8 @@
     "\n",
     "Launch the server in your terminal and wait for it to initialize.\n",
     "\n",
-    "Remember to add `--chat-template llama_3_vision` to specify the vision chat template, otherwise the server only supports text.\n",
+    "**Remember to add `--chat-template llama_3_vision` to specify the vision chat template, otherwise the server only supports text, and performance degradation may occur.**\n",
+    "\n",
     "We need to specify `--chat-template` for vision language models because the chat template provided in Hugging Face tokenizer only supports text."
    ]
   },

docs/backend/server_arguments.md

@@ -56,6 +56,9 @@ Please consult the documentation below to learn more about the parameters you ma
 * `json_model_override_args`: Override model config with the provided JSON.
 * `delete_ckpt_after_loading`: Delete the model checkpoint after loading the model.
 
+> [!IMPORTANT]
+> **Make sure the correct `chat_template` is passed, or performance degradation may occur.**
+
 ## Serving: HTTP & API
 
 ### HTTP Server configuration

python/sglang/lang/chat_template.py

@@ -353,7 +353,6 @@ register_chat_template(
     )
 )
 
-
 register_chat_template(
     ChatTemplate(
         name="deepseek-v3",
@@ -428,12 +427,13 @@ def match_chat_ml(model_path: str):
     if "tinyllama" in model_path:
         return get_chat_template("chatml")
     # Now the suffix for qwen2 chat model is "instruct"
-    if (
-        "qwen" in model_path
-        and ("chat" in model_path or "instruct" in model_path)
-        and ("llava" not in model_path)
-    ):
-        return get_chat_template("qwen")
+    if "qwen" in model_path:
+        if "vl" in model_path:
+            return get_chat_template("qwen2-vl")
+        if ("chat" in model_path or "instruct" in model_path) and (
+            "llava" not in model_path
+        ):
+            return get_chat_template("qwen")
     if (
         "llava-v1.6-34b" in model_path
         or "llava-v1.6-yi-34b" in model_path
@@ -459,6 +459,13 @@ def match_gemma_it(model_path: str):
         return get_chat_template("gemma-it")
 
 
+@register_chat_template_matching_function
+def match_openbmb_minicpm(model_path: str):
+    model_path = model_path.lower()
+    if "minicpm" in model_path:
+        return get_chat_template("minicpmv")
+
+
 @register_chat_template_matching_function
 def match_c4ai_command_r(model_path: str):
     model_path = model_path.lower()

python/sglang/srt/entrypoints/engine.py

@@ -438,7 +438,9 @@ def _launch_subprocesses(server_args: ServerArgs) -> Tuple[TokenizerManager, Dic
     # Launch tokenizer process
     tokenizer_manager = TokenizerManager(server_args, port_args)
     if server_args.chat_template:
-        load_chat_template_for_openai_api(tokenizer_manager, server_args.chat_template)
+        load_chat_template_for_openai_api(
+            tokenizer_manager, server_args.chat_template, server_args.model_path
+        )
 
     # Wait for the model to finish loading
     scheduler_infos = []

python/sglang/srt/models/llava.py

@@ -449,7 +449,8 @@ class LlavaBaseForCausalLM(nn.Module):
         projector_weights = {
             "model.mm_projector.0": "multi_modal_projector.linear_1",
             "model.mm_projector.2": "multi_modal_projector.linear_2",
-            "model.vision_tower.vision_tower": "vision_tower",  # Update the vision tower weights if we find them in the checkpoint (it may be finetuned).
+            "model.vision_tower.vision_tower": "vision_tower",
+            # Update the vision tower weights if we find them in the checkpoint (it may be finetuned).
             "model.image_newline": "language_model.model.image_newline",
         }
         params_dict = dict(self.named_parameters())

python/sglang/srt/openai_api/adapter.py

@@ -20,12 +20,14 @@ import os
 import time
 import uuid
 from http import HTTPStatus
-from typing import Dict, List, Optional
+from typing import Dict, List
 
 from fastapi import HTTPException, Request, UploadFile
 from fastapi.responses import ORJSONResponse, StreamingResponse
 from pydantic import ValidationError
 
+from sglang.lang.chat_template import get_chat_template_by_model_path
+
 try:
     from outlines.fsm.json_schema import convert_json_schema_to_str
 except ImportError:
@@ -92,7 +94,6 @@ file_id_response: Dict[str, FileResponse] = {}
 # map file id to file path in SGLang backend
 file_id_storage: Dict[str, str] = {}
 
-
 # backend storage directory
 storage_dir = None
 
@@ -116,12 +117,13 @@ def create_streaming_error_response(
     return json_str
 
 
-def load_chat_template_for_openai_api(tokenizer_manager, chat_template_arg):
+def load_chat_template_for_openai_api(tokenizer_manager, chat_template_arg, model_path):
     global chat_template_name
 
     logger.info(
         f"Use chat template for the OpenAI-compatible API server: {chat_template_arg}"
     )
+
     if not chat_template_exists(chat_template_arg):
         if not os.path.exists(chat_template_arg):
             raise RuntimeError(
@@ -163,6 +165,18 @@ def load_chat_template_for_openai_api(tokenizer_manager, chat_template_arg):
     else:
         chat_template_name = chat_template_arg
 
+    # check chat-template
+    chat_template = get_chat_template_by_model_path(model_path)
+    if chat_template is not None:
+        official_chat_template = chat_template.name
+        used_chat_template = chat_template_name
+        if official_chat_template != used_chat_template:
+            logger.warning(
+                f"Using a chat_template: '{used_chat_template}', "
+                f"which is different from official chat template: '{official_chat_template}', "
+                f"This discrepancy may lead to performance degradation."
+            )
+
 
 async def v1_files_create(file: UploadFile, purpose: str, file_storage_pth: str = None):
     try: