[Chore] Deprecate SupportsMultiModal.merge_by_field_config

vllm/model_executor/models/aria.py

@@ -499,8 +499,6 @@ class AriaForConditionalGeneration(nn.Module, SupportsMultiModal):
     model to perform tasks that involve both image and text inputs.
     """
 
-    merge_by_field_config = True
-
     hf_to_vllm_mapper = WeightsMapper(
         orig_to_new_prefix={
             # mapping for new names in checkpoint saved after transformers v4.52

vllm/model_executor/models/aya_vision.py

@@ -318,8 +318,6 @@ def _get_layer_index(feature_layer_index: int, num_hidden_layers: int) -> int:
     dummy_inputs=AyaVisionDummyInputsBuilder,
 )
 class AyaVisionForConditionalGeneration(nn.Module, SupportsMultiModal, SupportsPP):
-    merge_by_field_config = True
-
     hf_to_vllm_mapper = WeightsMapper(
         orig_to_new_prefix={
             # mapping for new names in checkpoint saved after transformers v4.52

vllm/model_executor/models/blip2.py

@@ -523,8 +523,6 @@ def _get_prompt_updates(
 class Blip2ForConditionalGeneration(
     nn.Module, SupportsMultiModal, SupportsPP, SupportsQuant
 ):
-    merge_by_field_config = True
-
     @classmethod
     def get_placeholder_str(cls, modality: str, i: int) -> str | None:
         if modality.startswith("image"):

vllm/model_executor/models/chameleon.py

@@ -918,8 +918,6 @@ def forward(
 class ChameleonForConditionalGeneration(
     nn.Module, SupportsMultiModal, SupportsPP, SupportsQuant
 ):
-    merge_by_field_config = True
-
     packed_modules_mapping = {
         "qkv_proj": ["q_proj", "k_proj", "v_proj"],
         "gate_up_proj": ["gate_proj", "up_proj"],

vllm/model_executor/models/clip.py

@@ -784,7 +784,6 @@ class CLIPEmbeddingModel(nn.Module, SupportsMultiModal, SupportsQuant):
     is_pooling_model = True
 
     packed_modules_mapping = {"qkv_proj": ["q_proj", "k_proj", "v_proj"]}
-    merge_by_field_config = True
 
     @classmethod
     def get_placeholder_str(cls, modality: str, i: int) -> str | None:

vllm/model_executor/models/cohere2_vision.py

@@ -331,8 +331,6 @@ def get_replacement(item_idx: int):
     dummy_inputs=Cohere2VisionDummyInputsBuilder,
 )
 class Cohere2VisionForConditionalGeneration(nn.Module, SupportsMultiModal, SupportsPP):
-    merge_by_field_config = True
-
     hf_to_vllm_mapper = WeightsMapper(
         orig_to_new_prefix={
             "model.vision_tower.": "vision_tower.",

vllm/model_executor/models/deepseek_ocr.py

@@ -344,8 +344,6 @@ def get_replacement_deepseek_vl2(item_idx: int):
     dummy_inputs=DeepseekOCRDummyInputsBuilder,
 )
 class DeepseekOCRForCausalLM(nn.Module, SupportsMultiModal, SupportsPP):
-    merge_by_field_config = True
-
     hf_to_vllm_mapper = WeightsMapper(
         orig_to_new_prefix={
             # map prefix for language backbone

vllm/model_executor/models/deepseek_vl2.py

@@ -344,8 +344,6 @@ def _cached_apply_hf_processor(
     dummy_inputs=DeepseekVL2DummyInputsBuilder,
 )
 class DeepseekVLV2ForCausalLM(nn.Module, SupportsMultiModal, SupportsPP):
-    merge_by_field_config = True
-
     hf_to_vllm_mapper = WeightsMapper(
         orig_to_new_prefix={
             "language.": "language_model.",

vllm/model_executor/models/dots_ocr.py

@@ -690,8 +690,6 @@ def forward(
     dummy_inputs=DotsOCRDummyInputsBuilder,
 )
 class DotsOCRForCausalLM(nn.Module, SupportsMultiModal, SupportsPP, SupportsLoRA):
-    merge_by_field_config = True
-
     hf_to_vllm_mapper = WeightsMapper(
         orig_to_new_substr={
             ".attn.qkv_proj.": ".attn.qkv.",

vllm/model_executor/models/ernie45_vl.py

@@ -1254,8 +1254,6 @@ def get_dummy_mm_data(
 class Ernie4_5_VLMoeForConditionalGeneration(
     nn.Module, SupportsMultiModal, SupportsLoRA, SupportsPP, SupportsMRoPE
 ):
-    merge_by_field_config = True
-
     packed_modules_mapping = {
         "qkv_proj": [
             "q_proj",

vllm/model_executor/models/fuyu.py

@@ -260,8 +260,6 @@ def get_replacement_fuyu(item_idx: int):
     dummy_inputs=FuyuDummyInputsBuilder,
 )
 class FuyuForCausalLM(nn.Module, SupportsMultiModal, SupportsPP):
-    merge_by_field_config = True
-
     hf_to_vllm_mapper = WeightsMapper(
         orig_to_new_prefix={
             "model.vision_embed_tokens.": "vision_embed_tokens.",

vllm/model_executor/models/gemma3_mm.py

@@ -483,8 +483,6 @@ def forward(self, vision_outputs: torch.Tensor):
 class Gemma3ForConditionalGeneration(
     nn.Module, SupportsMultiModal, SupportsPP, SupportsLoRA
 ):
-    merge_by_field_config = True
-
     packed_modules_mapping = {
         "qkv_proj": [
             "q_proj",

vllm/model_executor/models/gemma3n_mm.py

@@ -463,7 +463,6 @@ def forward(
 class Gemma3nForConditionalGeneration(
     nn.Module, SupportsMultiModal, SupportsTranscription
 ):
-    merge_by_field_config = True
     supported_languages = ISO639_1_SUPPORTED_LANGS
 
     packed_modules_mapping = {

vllm/model_executor/models/glm4_1v.py

@@ -1424,8 +1424,6 @@ def get_video_replacement_glm4v(item_idx: int):
 class Glm4vForConditionalGeneration(
     nn.Module, SupportsMultiModal, SupportsLoRA, SupportsPP, SupportsMRoPE
 ):
-    merge_by_field_config = True
-
     packed_modules_mapping = {
         "qkv_proj": [
             "q_proj",

vllm/model_executor/models/glm4v.py

@@ -561,8 +561,6 @@ def get_replacement(item_idx: int):
 class GLM4VForCausalLM(
     ChatGLMBaseModel, SupportsMultiModal, SupportsLoRA, SupportsPP, SupportsMRoPE
 ):
-    merge_by_field_config = True
-
     packed_modules_mapping = {
         "query_key_value": ["query_key_value"],
         "dense_h_to_4h": ["dense_h_to_4h"],

vllm/model_executor/models/granite_speech.py

@@ -564,7 +564,6 @@ class GraniteSpeechForConditionalGeneration(
     SupportsLoRA,
     SupportsTranscription,
 ):
-    merge_by_field_config = True
     supported_languages = ISO639_1_SUPPORTED_LANGS
 
     packed_modules_mapping = {

vllm/model_executor/models/hunyuan_vision.py

@@ -786,7 +786,6 @@ class HunYuanVLForConditionalGeneration(
     SupportsQuant,
     SupportsXDRoPE,
 ):
-    merge_by_field_config = True
     multimodal_cpu_fields = {"image_grid_thw"}
 
     # To ensure correct weight loading and mapping.

vllm/model_executor/models/hyperclovax_vision.py

@@ -592,8 +592,6 @@ def build_mlp(
     dummy_inputs=HCXVisionDummyInputsBuilder,
 )
 class HCXVisionForCausalLM(nn.Module, SupportsMultiModal, SupportsPP):
-    merge_by_field_config = True
-
     packed_modules_mapping = {
         "qkv_proj": ["q_proj", "k_proj", "v_proj"],
         "gate_up_proj": ["gate_proj", "up_proj"],

vllm/model_executor/models/idefics3.py

@@ -576,8 +576,6 @@ def forward(
     dummy_inputs=Idefics3DummyInputsBuilder,
 )
 class Idefics3ForConditionalGeneration(nn.Module, SupportsMultiModal, SupportsLoRA):
-    merge_by_field_config = True
-
     packed_modules_mapping = {
         "qkv_proj": [
             "q_proj",

vllm/model_executor/models/interfaces.py

@@ -78,9 +78,9 @@ class SupportsMultiModal(Protocol):
     `multimodal_config.mm_encoder_tp_mode="data"`.
     """
 
-    merge_by_field_config: ClassVar[bool] = True
+    merge_by_field_config: ClassVar[bool | None] = None
     """
-    A flag that indicates which implementation of
+    [DEPRECATED] A flag that indicates which implementation of
     `vllm.multimodal.utils.group_mm_kwargs_by_modality` to use.
     """
 
@@ -260,7 +260,26 @@ def supports_multimodal(model: object) -> TypeIs[SupportsMultiModal]: ...
 def supports_multimodal(
     model: type[object] | object,
 ) -> TypeIs[type[SupportsMultiModal]] | TypeIs[SupportsMultiModal]:
-    return getattr(model, "supports_multimodal", False)
+    res = getattr(model, "supports_multimodal", False)
+
+    if res:
+        # We can remove this starting from v0.14
+        merge_by_field_config = getattr(model, "merge_by_field_config", None)
+        if merge_by_field_config is False:
+            raise ValueError(
+                "`merge_by_field_config=False` is no longer effective, "
+                "please update your model to consider the new batching logic "
+                "in `group_mm_kwargs_by_modality` (refer to "
+                "https://github.com/vllm-project/vllm/issues/26149), "
+                "and then remove the override from your model."
+            )
+        if merge_by_field_config is True:
+            logger.warning_once(
+                "`merge_by_field_config=True` is redundant, "
+                "please remove the override from your model."
+            )
+
+    return res
 
 
 def supports_multimodal_raw_input_only(model: type[object] | object) -> bool:

vllm/model_executor/models/interns1.py

@@ -509,8 +509,6 @@ def get_replacement_interns1_video(item_idx: int):
 class InternS1ForConditionalGeneration(
     nn.Module, SupportsMultiModal, SupportsPP, SupportsLoRA
 ):
-    merge_by_field_config = True
-
     # To ensure correct weight loading and mapping.
     hf_to_vllm_mapper = WeightsMapper(
         orig_to_new_prefix={

vllm/model_executor/models/internvl.py

@@ -1074,8 +1074,6 @@ def get_video_replacement_internvl(item_idx: int):
     dummy_inputs=InternVLDummyInputsBuilder,
 )
 class InternVLChatModel(nn.Module, SupportsMultiModal, SupportsPP, SupportsLoRA):
-    merge_by_field_config = True
-
     supports_encoder_tp_data = True
 
     @classmethod

vllm/model_executor/models/keye.py

@@ -1292,8 +1292,6 @@ def _get_mm_fields_config(
 
 
 class BaseKeyeModule(nn.Module):
-    merge_by_field_config = True
-
     packed_modules_mapping = {
         "qkv_proj": [
             "q_proj",

vllm/model_executor/models/kimi_vl.py

@@ -298,8 +298,6 @@ def get_replacement(item_idx: int):
     dummy_inputs=KimiVLDummyInputsBuilder,
 )
 class KimiVLForConditionalGeneration(nn.Module, SupportsMultiModal, SupportsPP):
-    merge_by_field_config = True
-
     supports_encoder_tp_data = True
 
     @classmethod

vllm/model_executor/models/llava.py

@@ -506,8 +506,6 @@ def init_vision_tower_for_llava(
     dummy_inputs=LlavaDummyInputsBuilder,
 )
 class LlavaForConditionalGeneration(nn.Module, SupportsMultiModal, SupportsPP):
-    merge_by_field_config = True
-
     packed_modules_mapping = {
         "qkv_proj": ["q_proj", "k_proj", "v_proj"],
         "gate_up_proj": ["gate_proj", "up_proj"],

vllm/model_executor/models/llava_next.py

@@ -223,8 +223,6 @@ def _get_mm_fields_config(
     dummy_inputs=LlavaDummyInputsBuilder,
 )
 class LlavaNextForConditionalGeneration(nn.Module, SupportsMultiModal, SupportsPP):
-    merge_by_field_config = True
-
     hf_to_vllm_mapper = WeightsMapper(
         orig_to_new_prefix={
             # mapping for new names in checkpoint saved after transformers v4.52

vllm/model_executor/models/llava_next_video.py

@@ -299,8 +299,6 @@ def forward(self, image_features: torch.Tensor) -> torch.Tensor:
     dummy_inputs=LlavaNextVideoDummyInputsBuilder,
 )
 class LlavaNextVideoForConditionalGeneration(nn.Module, SupportsMultiModal, SupportsPP):
-    merge_by_field_config = True
-
     hf_to_vllm_mapper = WeightsMapper(
         orig_to_new_prefix={
             # mapping for new names in checkpoint saved after transformers v4.52

vllm/model_executor/models/llava_onevision.py

@@ -479,8 +479,6 @@ def forward(self, image_features: torch.Tensor) -> torch.Tensor:
     dummy_inputs=LlavaOnevisionDummyInputsBuilder,
 )
 class LlavaOnevisionForConditionalGeneration(nn.Module, SupportsMultiModal, SupportsPP):
-    merge_by_field_config = True
-
     hf_to_vllm_mapper = WeightsMapper(
         orig_to_new_prefix={
             # mapping for new names in checkpoint saved after transformers v4.52

vllm/model_executor/models/midashenglm.py

@@ -683,8 +683,6 @@ def get_replacement_midashenglm(item_idx: int):
     dummy_inputs=MiDashengLMDummyInputsBuilder,
 )
 class MiDashengLMModel(nn.Module, SupportsMultiModal, SupportsPP):
-    merge_by_field_config = True
-
     packed_modules_mapping = {
         "qkv_proj": [
             "q_proj",

vllm/model_executor/models/minicpmv.py

@@ -1003,8 +1003,6 @@ class MiniCPMVBaseModel(nn.Module, SupportsMultiModal, SupportsPP):
     instantiated.
     """
 
-    merge_by_field_config = True
-
     supports_encoder_tp_data = True
 
     @classmethod

vllm/model_executor/models/minimax_vl_01.py

@@ -179,8 +179,6 @@ def _get_mm_fields_config(
     dummy_inputs=MiniMaxVL01DummyInputsBuilder,
 )
 class MiniMaxVL01ForConditionalGeneration(nn.Module, SupportsMultiModal, SupportsPP):
-    merge_by_field_config = True
-
     packed_modules_mapping = {
         "qkv_proj": ["q_proj", "k_proj", "v_proj"],
         "gate_up_proj": ["gate_proj", "up_proj"],

vllm/model_executor/models/mistral3.py

@@ -423,8 +423,6 @@ def init_vision_tower_for_llava(
 class Mistral3ForConditionalGeneration(
     nn.Module, SupportsLoRA, SupportsMultiModal, SupportsPP
 ):
-    merge_by_field_config = True
-
     packed_modules_mapping = {
         "qkv_proj": ["q_proj", "k_proj", "v_proj"],
         "gate_up_proj": ["gate_proj", "up_proj"],

vllm/model_executor/models/mllama4.py

@@ -741,8 +741,6 @@ class Llama4ForConditionalGeneration(
     SupportsEagle3,
     SupportsLoRA,
 ):
-    merge_by_field_config = True
-
     packed_modules_mapping = {
         "qkv_proj": ["q_proj", "k_proj", "v_proj"],
         "gate_up_proj": ["gate_proj", "up_proj"],

vllm/model_executor/models/molmo.py

@@ -1354,8 +1354,6 @@ def get_insertion_molmo(item_idx: int):
 class MolmoForCausalLM(
     nn.Module, SupportsMultiModal, SupportsPP, SupportsLoRA, SupportsQuant
 ):
-    merge_by_field_config = True
-
     hf_to_vllm_mapper = WeightsMapper(
         orig_to_new_substr={
             # vision backbone mapping

vllm/model_executor/models/nano_nemotron_vl.py

@@ -1116,8 +1116,6 @@ def get_dummy_mm_data(
 class NemotronH_Nano_VL_V2(
     nn.Module, HasInnerState, IsHybrid, SupportsMultiModal, SupportsMultiModalPruning
 ):
-    merge_by_field_config = True
-
     @classmethod
     def get_placeholder_str(cls, modality: str, i: int) -> str | None:
         if modality.startswith("image"):

vllm/model_executor/models/nemotron_vl.py

@@ -358,8 +358,6 @@ def get_image_processor(self, **kwargs: object):
     dummy_inputs=BaseInternVLDummyInputsBuilder[NemotronVLProcessingInfo],
 )
 class LlamaNemotronVLChatModel(nn.Module, SupportsMultiModal, SupportsPP, SupportsLoRA):
-    merge_by_field_config = True
-
     @classmethod
     def get_placeholder_str(cls, modality: str, i: int) -> str | None:
         if modality.startswith("image"):

vllm/model_executor/models/opencua.py

@@ -201,7 +201,6 @@ def get_dummy_text(self, mm_counts: Mapping[str, int]) -> str:
     dummy_inputs=OpenCUADummyInputsBuilder,
 )
 class OpenCUAForConditionalGeneration(Qwen2_5_VLForConditionalGeneration):
-    merge_by_field_config = True
     multimodal_cpu_fields = {"image_grid_thw"}
 
     packed_modules_mapping = {

vllm/model_executor/models/ovis.py

@@ -414,8 +414,6 @@ def get_replacement_ovis(item_idx: int):
     dummy_inputs=OvisDummyInputsBuilder,
 )
 class Ovis(nn.Module, SupportsMultiModal, SupportsPP):
-    merge_by_field_config = True
-
     @classmethod
     def get_placeholder_str(cls, modality: str, i: int) -> str | None:
         if modality.startswith("image"):

vllm/model_executor/models/ovis2_5.py

@@ -456,8 +456,6 @@ def get_replacement_ovis(item_idx, modality: str):
     dummy_inputs=Ovis2_5DummyInputsBuilder,
 )
 class Ovis2_5(nn.Module, SupportsMultiModal, SupportsPP):
-    merge_by_field_config = True
-
     def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
         super().__init__()
         config = vllm_config.model_config.hf_config

vllm/model_executor/models/paddleocr_vl.py

@@ -1103,8 +1103,6 @@ def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
     dummy_inputs=PaddleOCRVLDummyInputsBuilder,
 )
 class PaddleOCRVLForConditionalGeneration(nn.Module, SupportsMultiModal, SupportsMRoPE):
-    merge_by_field_config = True
-
     hf_to_vllm_mapper = WeightsMapper(
         orig_to_new_prefix={
             "model.": "language_model.model.",

vllm/model_executor/models/paligemma.py

@@ -251,8 +251,6 @@ def apply(
     dummy_inputs=PaliGemmaDummyInputsBuilder,
 )
 class PaliGemmaForConditionalGeneration(nn.Module, SupportsMultiModal, SupportsPP):
-    merge_by_field_config = True
-
     packed_modules_mapping = {
         "qkv_proj": [
             "q_proj",