[Core] Move multimodal placeholder from chat utils to model definition (vllm-project#20355)

Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>

Author: DarkLight1337

docs/contributing/model/multimodal.md

@@ -10,6 +10,22 @@ This document walks you through the steps to extend a basic model so that it acc
 It is assumed that you have already implemented the model in vLLM according to [these steps][new-model-basic].
 Further update the model as follows:
 
+- Implement [get_placeholder_str][vllm.model_executor.models.interfaces.SupportsMultiModal.get_placeholder_str] to define the placeholder string which is used to represent the multi-modal item in the text prompt. This should be consistent with the chat template of the model.
+
+    ??? Code
+
+        ```python
+        class YourModelForImage2Seq(nn.Module):
+            ...
+
+            @classmethod
+            def get_placeholder_str(cls, modality: str, i: int) -> Optional[str]:
+                if modality.startswith("image"):
+                    return "<image>"
+
+                raise ValueError("Only image modality is supported")
+        ```
+
 - Reserve a keyword parameter in [forward][torch.nn.Module.forward] for each input tensor that corresponds to a multi-modal input, as shown in the following example:
 
   ```diff

tests/async_engine/test_async_llm_engine.py

@@ -33,7 +33,6 @@ class RequestOutput:
 class MockModelConfig:
     use_async_output_proc = True
     media_io_kwargs: dict[str, dict[str, Any]] = field(default_factory=dict)
-    mm_placeholder_str_override: dict[str, str] = field(default_factory=dict)
 
 
 class MockEngine:

tests/engine/test_arg_utils.py

@@ -263,26 +263,6 @@ def test_media_io_kwargs_parser(arg, expected):
     assert args.media_io_kwargs == expected
 
 
-@pytest.mark.parametrize(("arg", "expected"), [
-    (None, dict()),
-    ('{"video":"<|video_placeholder|>"}', {
-        "video": "<|video_placeholder|>"
-    }),
-    ('{"video":"<|video_placeholder|>", "image": "<|image_placeholder|>"}', {
-        "video": "<|video_placeholder|>",
-        "image": "<|image_placeholder|>"
-    }),
-])
-def test_mm_placeholder_str_override_parser(arg, expected):
-    parser = EngineArgs.add_cli_args(FlexibleArgumentParser())
-    if arg is None:
-        args = parser.parse_args([])
-    else:
-        args = parser.parse_args(["--mm-placeholder-str-override", arg])
-
-    assert args.mm_placeholder_str_override == expected
-
-
 def test_compilation_config():
     parser = EngineArgs.add_cli_args(FlexibleArgumentParser())
 

tests/entrypoints/openai/test_serving_chat.py

@@ -41,7 +41,6 @@ class MockModelConfig:
     encoder_config = None
     generation_config: str = "auto"
     media_io_kwargs: dict[str, dict[str, Any]] = field(default_factory=dict)
-    mm_placeholder_str_override: dict[str, str] = field(default_factory=dict)
 
     def get_diff_sampling_param(self):
         return self.diff_sampling_param or {}

vllm/config.py

@@ -350,8 +350,6 @@ class ModelConfig:
     """Additional args passed to process media inputs, keyed by modalities. 
     For example, to set num_frames for video, set 
     `--media-io-kwargs '{"video": {"num_frames": 40} }'` """
-    mm_placeholder_str_override: dict[str, str] = field(default_factory=dict)
-    """Optionally override placeholder string for given modalities."""
     use_async_output_proc: bool = True
     """Whether to use async output processor."""
     config_format: Union[str, ConfigFormat] = ConfigFormat.AUTO.value
@@ -661,7 +659,7 @@ def architecture(self) -> str:
         return self._architecture
 
     @property
-    def model_info(self) -> dict[str, Any]:
+    def model_info(self):
         return self._model_info
 
     def maybe_pull_model_tokenizer_for_s3(self, model: str,
@@ -701,7 +699,6 @@ def _init_multimodal_config(self) -> Optional["MultiModalConfig"]:
             return MultiModalConfig(
                 limit_per_prompt=self.limit_mm_per_prompt,
                 media_io_kwargs=self.media_io_kwargs,
-                mm_placeholder_str_override=self.mm_placeholder_str_override,
                 mm_processor_kwargs=self.mm_processor_kwargs,
                 disable_mm_preprocessor_cache=self.
                 disable_mm_preprocessor_cache)
@@ -3096,9 +3093,6 @@ class MultiModalConfig:
     For example, to set num_frames for video, set 
     `--media-io-kwargs '{"video": {"num_frames": 40} }'` """
 
-    mm_placeholder_str_override: dict[str, str] = field(default_factory=dict)
-    """Optionally override placeholder string for given modalities."""
-
     mm_processor_kwargs: Optional[dict[str, object]] = None
     """
     Overrides for the multi-modal processor obtained from

vllm/engine/arg_utils.py

@@ -373,8 +373,6 @@ class EngineArgs:
     media_io_kwargs: dict[str, dict[str,
                                     Any]] = get_field(MultiModalConfig,
                                                       "media_io_kwargs")
-    mm_placeholder_str_override: dict[str, str] = \
-        get_field(MultiModalConfig, "mm_placeholder_str_override")
     mm_processor_kwargs: Optional[Dict[str, Any]] = \
         MultiModalConfig.mm_processor_kwargs
     disable_mm_preprocessor_cache: bool = \
@@ -759,9 +757,6 @@ def add_cli_args(parser: FlexibleArgumentParser) -> FlexibleArgumentParser:
                                       **multimodal_kwargs["limit_per_prompt"])
         multimodal_group.add_argument("--media-io-kwargs",
                                       **multimodal_kwargs["media_io_kwargs"])
-        multimodal_group.add_argument(
-            "--mm-placeholder-str-override",
-            **multimodal_kwargs["mm_placeholder_str_override"])
         multimodal_group.add_argument(
             "--mm-processor-kwargs",
             **multimodal_kwargs["mm_processor_kwargs"])
@@ -987,7 +982,6 @@ def create_model_config(self) -> ModelConfig:
             served_model_name=self.served_model_name,
             limit_mm_per_prompt=self.limit_mm_per_prompt,
             media_io_kwargs=self.media_io_kwargs,
-            mm_placeholder_str_override=self.mm_placeholder_str_override,
             use_async_output_proc=not self.disable_async_output_proc,
             config_format=self.config_format,
             mm_processor_kwargs=self.mm_processor_kwargs,

vllm/entrypoints/chat_utils.py

@@ -6,7 +6,7 @@
 from abc import ABC, abstractmethod
 from collections import defaultdict, deque
 from collections.abc import Awaitable, Iterable
-from functools import cache, lru_cache, partial
+from functools import cached_property, lru_cache, partial
 from pathlib import Path
 from typing import (Any, Callable, Generic, Literal, Optional, TypeVar, Union,
                     cast)
@@ -37,6 +37,8 @@
 
 from vllm.config import ModelConfig
 from vllm.logger import init_logger
+from vllm.model_executor.model_loader import get_model_cls
+from vllm.model_executor.models import SupportsMultiModal
 from vllm.multimodal import MULTIMODAL_REGISTRY, MultiModalDataDict
 from vllm.multimodal.utils import MediaConnector
 # yapf: disable
@@ -492,6 +494,10 @@ def __init__(self, model_config: ModelConfig, tokenizer: AnyTokenizer):
     def model_config(self) -> ModelConfig:
         return self._model_config
 
+    @cached_property
+    def model_cls(self):
+        return get_model_cls(self.model_config)
+
     @property
     def allowed_local_media_path(self):
         return self._model_config.allowed_local_media_path
@@ -500,96 +506,14 @@ def allowed_local_media_path(self):
     def mm_registry(self):
         return MULTIMODAL_REGISTRY
 
-    @staticmethod
-    @cache
-    def _cached_token_str(tokenizer: AnyTokenizer, token_index: int) -> str:
-        return tokenizer.decode(token_index)
-
-    def _placeholder_str(self, modality: ModalityStr,
-                         current_count: int) -> Optional[str]:
-        if modality in self._model_config.mm_placeholder_str_override:
-            return self._model_config.mm_placeholder_str_override[modality]
-
-        # TODO: Let user specify how to insert image tokens into prompt
-        # (similar to chat template)
-        hf_config = self._model_config.hf_config
-        model_type = hf_config.model_type
-
-        if modality in ("image", "image_embeds"):
-            if model_type == "chatglm":
-                return "<|begin_of_image|><|endoftext|><|end_of_image|>"
-            if model_type == "glm4v":
-                return "<|begin_of_image|><|image|><|end_of_image|>"
-            if model_type in ("phi3_v", "phi4mm"):
-                return f"<|image_{current_count}|>"
-            if model_type in ("minicpmo", "minicpmv"):
-                return "(<image>./</image>)"
-            if model_type in ("blip-2", "florence2", "fuyu", "paligemma",
-                              "pixtral", "mistral3"):
-                # These models do not use image tokens in the prompt
-                return None
-            if model_type == "qwen":
-                return f"Picture {current_count}: <img></img>"
-            if model_type.startswith("llava"):
-                return self._cached_token_str(self._tokenizer,
-                                              hf_config.image_token_index)
-
-            if model_type in ("aya_vision", "chameleon", "deepseek_vl_v2",
-                              "internvl_chat", "ovis", "skywork_chat",
-                              "NVLM_D", "h2ovl_chat", "idefics3", "smolvlm"):
-                return "<image>"
-            if model_type in ("mllama", "llama4"):
-                return "<|image|>"
-            if model_type in ("qwen2_vl", "qwen2_5_vl", "keye", "Keye"):
-                return "<|vision_start|><|image_pad|><|vision_end|>"
-            if model_type == "qwen2_5_omni":
-                return "<|vision_start|><|IMAGE|><|vision_end|>"
-            if model_type == "molmo":
-                return ""
-            if model_type == "aria":
-                return "<|fim_prefix|><|img|><|fim_suffix|>"
-            if model_type == "gemma3":
-                return "<start_of_image>"
-            if model_type == "kimi_vl":
-                return "<|media_start|>image<|media_content|><|media_pad|><|media_end|>" # noqa: E501
-
-            raise TypeError(f"Unknown {modality} model type: {model_type}")
-        elif modality == "audio":
-            if model_type in ("ultravox", "granite_speech"):
-                return "<|audio|>"
-            if model_type == "phi4mm":
-                return f"<|audio_{current_count}|>"
-            if model_type in ("qwen2_audio", "qwen2_5_omni"):
-                return (f"Audio {current_count}: "
-                        f"<|audio_bos|><|AUDIO|><|audio_eos|>")
-            if model_type == "minicpmo":
-                return "(<audio>./</audio>)"
-            raise TypeError(f"Unknown model type: {model_type}")
-        elif modality == "video":
-            if model_type == "internvl_chat":
-                return "<video>"
-            if model_type == "glm4v":
-                return "<|begin_of_video|><|video|><|end_of_video|>"
-            if model_type in ("qwen2_vl", "qwen2_5_vl", "keye", "Keye"):
-                return "<|vision_start|><|video_pad|><|vision_end|>"
-            if model_type == "qwen2_5_omni":
-                return "<|vision_start|><|VIDEO|><|vision_end|>"
-            if model_type in ("minicpmo", "minicpmv"):
-                return "(<video>./</video>)"
-            if model_type.startswith("llava"):
-                return self._cached_token_str(self._tokenizer,
-                                              hf_config.video_token_index)
-            raise TypeError(f"Unknown {modality} model type: {model_type}")
-        else:
-            raise TypeError(f"Unknown modality: {modality}")
-
     def add(self, modality: ModalityStr, item: _T) -> Optional[str]:
         """
         Add a multi-modal item to the current prompt and returns the
         placeholder string to use, if any.
         """
         mm_registry = self.mm_registry
         model_config = self.model_config
+        model_cls = cast(SupportsMultiModal, self.model_cls)
 
         input_modality = modality.replace("_embeds", "")
 
@@ -614,7 +538,7 @@ def add(self, modality: ModalityStr, item: _T) -> Optional[str]:
 
         self._items_by_modality[modality].append(item)
 
-        return self._placeholder_str(modality, current_count)
+        return model_cls.get_placeholder_str(modality, current_count)
 
     @abstractmethod
     def create_parser(self) -> "BaseMultiModalContentParser":

vllm/entrypoints/openai/speech_to_text.py

@@ -5,6 +5,7 @@
 import math
 import time
 from collections.abc import AsyncGenerator
+from functools import cached_property
 from math import ceil
 from typing import Callable, Literal, Optional, TypeVar, Union, cast
 
@@ -24,7 +25,8 @@
 from vllm.entrypoints.openai.serving_models import OpenAIServingModels
 from vllm.inputs.data import PromptType
 from vllm.logger import init_logger
-from vllm.model_executor.model_loader.utils import get_model_architecture
+from vllm.model_executor.model_loader import get_model_cls
+from vllm.model_executor.models import SupportsTranscription
 from vllm.outputs import RequestOutput
 from vllm.transformers_utils.processor import cached_get_processor
 from vllm.utils import PlaceholderModule
@@ -76,24 +78,29 @@ def __init__(
         self.model_sr = processor.feature_extractor.sampling_rate
         self.hop_length = processor.feature_extractor.hop_length
         self.task_type = task_type
-        self.model_cls, _ = get_model_architecture(model_config)
 
         if self.default_sampling_params:
             logger.info(
                 "Overwriting default completion sampling param with: %s",
                 self.default_sampling_params)
 
+    @cached_property
+    def model_cls(self):
+        return get_model_cls(self.model_config)
+
     async def _preprocess_speech_to_text(
         self,
         request: SpeechToTextRequest,
         audio_data: bytes,
     ) -> tuple[list[PromptType], float]:
+        model_cls = cast(SupportsTranscription, self.model_cls)
+
         # Validate request
         # TODO language should be optional and can be guessed.
         # For now we default to en. See
         # https://github.com/huggingface/transformers/blob/main/src/transformers/models/whisper/generation_whisper.py#L1520
         lang = request.language or "en"
-        self.model_cls.validate_language(lang)  # type: ignore[attr-defined]
+        model_cls.validate_language(lang)
 
         if len(audio_data) / 1024**2 > MAX_AUDIO_CLIP_FILESIZE_MB:
             raise ValueError("Maximum file size exceeded.")
@@ -117,9 +124,8 @@ async def _preprocess_speech_to_text(
                     },
                 },
                 "decoder_prompt":
-                self.model_cls.
-                get_decoder_prompt(  # type: ignore[attr-defined]
-                    lang, self.task_type, request.prompt)
+                model_cls.get_decoder_prompt(lang, self.task_type,
+                                             request.prompt)
             }
             prompts.append(cast(PromptType, prompt))
         return prompts, duration

vllm/model_executor/model_loader/__init__.py

@@ -18,7 +18,7 @@
     ShardedStateLoader)
 from vllm.model_executor.model_loader.tensorizer_loader import TensorizerLoader
 from vllm.model_executor.model_loader.utils import (
-    get_architecture_class_name, get_model_architecture)
+    get_architecture_class_name, get_model_architecture, get_model_cls)
 
 
 def get_model_loader(load_config: LoadConfig) -> BaseModelLoader:
@@ -65,6 +65,7 @@ def get_model(*,
     "get_model_loader",
     "get_architecture_class_name",
     "get_model_architecture",
+    "get_model_cls",
     "BaseModelLoader",
     "BitsAndBytesModelLoader",
     "GGUFModelLoader",

vllm/model_executor/model_loader/tensorizer.py

@@ -13,7 +13,7 @@
 from collections.abc import Generator
 from dataclasses import dataclass
 from functools import partial
-from typing import Any, BinaryIO, Optional, Union
+from typing import TYPE_CHECKING, Any, BinaryIO, Optional, Union
 
 import regex as re
 import torch
@@ -24,12 +24,14 @@
 import vllm.envs as envs
 from vllm.config import (ModelConfig, ParallelConfig, VllmConfig,
                          set_current_vllm_config)
-from vllm.engine.arg_utils import EngineArgs
 from vllm.logger import init_logger
 from vllm.model_executor.layers.vocab_parallel_embedding import (
     VocabParallelEmbedding)
 from vllm.utils import FlexibleArgumentParser, PlaceholderModule
 
+if TYPE_CHECKING:
+    from vllm.engine.arg_utils import EngineArgs
+
 try:
     from tensorizer import (DecryptionParams, EncryptionParams,
                             TensorDeserializer, TensorSerializer)
@@ -503,7 +505,7 @@ def serialize_vllm_model(
     return model
 
 
-def tensorize_vllm_model(engine_args: EngineArgs,
+def tensorize_vllm_model(engine_args: "EngineArgs",
                          tensorizer_config: TensorizerConfig,
                          generate_keyfile: bool = True):
     """Utility to load a model and then serialize it with Tensorizer