[Bugfix] Fix chat template loading (#15143)

Signed-off-by: DarkLight1337 <[email protected]>
Signed-off-by: Roger Wang <[email protected]>
Co-authored-by: chaunceyjiang <[email protected]>
Co-authored-by: Roger Wang <[email protected]>

Author: 3 people

tests/entrypoints/openai/test_chat_template.py

@@ -107,8 +107,10 @@ def test_get_gen_prompt(model, template, add_generation_prompt,
     # Call the function and get the result
     result = apply_hf_chat_template(
         tokenizer,
+        trust_remote_code=True,
         conversation=mock_request.messages,
         chat_template=mock_request.chat_template or template_content,
+        tools=None,
         add_generation_prompt=mock_request.add_generation_prompt,
         continue_final_message=mock_request.continue_final_message,
     )

tests/entrypoints/openai/test_video.py

@@ -87,7 +87,7 @@ async def test_single_chat_session_video(client: openai.AsyncOpenAI,
     choice = chat_completion.choices[0]
     assert choice.finish_reason == "length"
     assert chat_completion.usage == openai.types.CompletionUsage(
-        completion_tokens=10, prompt_tokens=6299, total_tokens=6309)
+        completion_tokens=10, prompt_tokens=6287, total_tokens=6297)
 
     message = choice.message
     message = chat_completion.choices[0].message
@@ -180,7 +180,7 @@ async def test_single_chat_session_video_base64encoded(
     choice = chat_completion.choices[0]
     assert choice.finish_reason == "length"
     assert chat_completion.usage == openai.types.CompletionUsage(
-        completion_tokens=10, prompt_tokens=6299, total_tokens=6309)
+        completion_tokens=10, prompt_tokens=6287, total_tokens=6297)
 
     message = choice.message
     message = chat_completion.choices[0].message

tests/entrypoints/test_chat_utils.py

@@ -4,10 +4,13 @@
 from typing import Optional
 
 import pytest
+from packaging.version import Version
+from transformers import __version__ as TRANSFORMERS_VERSION
 
 from vllm.assets.image import ImageAsset
 from vllm.config import ModelConfig
-from vllm.entrypoints.chat_utils import (_try_extract_ast, load_chat_template,
+from vllm.entrypoints.chat_utils import (_resolve_hf_chat_template,
+                                         _try_extract_ast, load_chat_template,
                                          parse_chat_messages,
                                          parse_chat_messages_futures,
                                          resolve_chat_template_content_format)
@@ -23,8 +26,10 @@
 PHI3V_MODEL_ID = "microsoft/Phi-3.5-vision-instruct"
 ULTRAVOX_MODEL_ID = "fixie-ai/ultravox-v0_5-llama-3_2-1b"
 QWEN2VL_MODEL_ID = "Qwen/Qwen2-VL-2B-Instruct"
+QWEN25VL_MODEL_ID = "Qwen/Qwen2.5-VL-3B-Instruct"
 MLLAMA_MODEL_ID = "meta-llama/Llama-3.2-11B-Vision-Instruct"
 LLAMA_GUARD_MODEL_ID = "meta-llama/Llama-Guard-3-1B"
+HERMES_MODEL_ID = "NousResearch/Hermes-3-Llama-3.1-8B"
 
 
 @pytest.fixture(scope="function")
@@ -703,25 +708,70 @@ def get_conversation(is_hf: bool):
 
     vllm_result = apply_hf_chat_template(
         tokenizer,
+        trust_remote_code=model_config.trust_remote_code,
         conversation=conversation,
         chat_template=None,
+        tools=None,
         add_generation_prompt=True,
     )
 
     assert hf_result == vllm_result
 
 
+@pytest.mark.parametrize(
+    "model",
+    [
+        QWEN2VL_MODEL_ID,  # tokenizer.chat_template is of type str
+        HERMES_MODEL_ID,  # tokenizer.chat_template is of type dict
+    ])
+@pytest.mark.parametrize("use_tools", [True, False])
+def test_resolve_hf_chat_template(sample_json_schema, model, use_tools):
+    """checks that chat_template is a dict type for HF models."""
+
+    # Build the tokenizer group and grab the underlying tokenizer
+    tokenizer_group = TokenizerGroup(
+        model,
+        enable_lora=False,
+        max_num_seqs=5,
+        max_input_length=None,
+    )
+    tokenizer = tokenizer_group.tokenizer
+
+    tools = [{
+        "type": "function",
+        "function": {
+            "name": "dummy_function_name",
+            "description": "This is a dummy function",
+            "parameters": sample_json_schema
+        }
+    }] if use_tools else None
+
+    # Test detecting the tokenizer's chat_template
+    chat_template = _resolve_hf_chat_template(
+        tokenizer,
+        chat_template=None,
+        tools=tools,
+        trust_remote_code=True,
+    )
+    assert isinstance(chat_template, str)
+
+
 # yapf: disable
 @pytest.mark.parametrize(
     ("model", "expected_format"),
     [(PHI3V_MODEL_ID, "string"),
      (QWEN2VL_MODEL_ID, "openai"),
+     (QWEN25VL_MODEL_ID, "openai"),
      (ULTRAVOX_MODEL_ID, "string"),
      (MLLAMA_MODEL_ID, "openai"),
      (LLAMA_GUARD_MODEL_ID, "openai")],
 )
 # yapf: enable
 def test_resolve_content_format_hf_defined(model, expected_format):
+    if model == QWEN25VL_MODEL_ID and Version(TRANSFORMERS_VERSION) < Version(
+            "4.49.0"):
+        pytest.skip("Qwen2.5-VL requires transformers>=4.49.0")
+
     tokenizer_group = TokenizerGroup(
         model,
         enable_lora=False,
@@ -730,7 +780,13 @@ def test_resolve_content_format_hf_defined(model, expected_format):
     )
     tokenizer = tokenizer_group.tokenizer
 
-    chat_template = tokenizer.chat_template
+    # Test detecting the tokenizer's chat_template
+    chat_template = _resolve_hf_chat_template(
+        tokenizer,
+        chat_template=None,
+        tools=None,
+        trust_remote_code=True,
+    )
     assert isinstance(chat_template, str)
 
     print("[TEXT]")
@@ -740,8 +796,10 @@ def test_resolve_content_format_hf_defined(model, expected_format):
 
     resolved_format = resolve_chat_template_content_format(
         None,  # Test detecting the tokenizer's chat_template
+        None,
         "auto",
         tokenizer,
+        trust_remote_code=True,
     )
 
     assert resolved_format == expected_format
@@ -791,8 +849,10 @@ def test_resolve_content_format_examples(template_path, expected_format):
 
     resolved_format = resolve_chat_template_content_format(
         chat_template,
+        None,
         "auto",
         dummy_tokenizer,
+        trust_remote_code=True,
     )
 
     assert resolved_format == expected_format

tests/tool_use/utils.py

@@ -39,7 +39,10 @@ def ensure_system_prompt(messages: list[dict[str, Any]],
 
 # universal args for all models go here. also good if you need to test locally
 # and change type or KV cache quantization or something.
-ARGS: list[str] = ["--enable-auto-tool-choice", "--max-model-len", "1024"]
+ARGS: list[str] = [
+    "--enable-auto-tool-choice", "--max-model-len", "1024", "--max-num-seqs",
+    "256"
+]
 
 CONFIGS: dict[str, ServerConfig] = {
     "hermes": {