[Chore] Deprecate `SupportsMultiModal.merge_by_field_config` (#30170)

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

vllm/model_executor/models/interns1.py

@@ -509,8 +509,6 @@ class InternS1MultiModalProcessor(BaseMultiModalProcessor[InternS1ProcessingInfo
 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 @@ class InternVLMultiModalProcessor(
     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 @@ class KeyeMultiModalProcessor(BaseMultiModalProcessor[KeyeProcessingInfo]):
 
 
 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 @@ class KimiVLMultiModalProcessor(BaseMultiModalProcessor[KimiVLProcessingInfo]):
     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 @@ class LlavaNextMultiModalProcessor(
     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 @@ class LlavaNextMultiModalProjector(nn.Module):
     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 @@ class LlavaOnevisionMultiModalProjector(nn.Module):
     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 @@ class MiDashengLMMultiModalProcessor(
     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 @@ class MiniMaxVL01MultiModalProcessor(
     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 @@ class MolmoMultiModalProcessor(BaseMultiModalProcessor[MolmoProcessingInfo]):
 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 @@ class NanoNemotronVLDummyInputsBuilder(
 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 @@ class NemotronVLProcessingInfo(BaseInternVLProcessingInfo):
     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 @@ class OpenCUADummyInputsBuilder(Qwen2VLDummyInputsBuilder):
     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 @@ class OvisMultiModalProcessor(BaseMultiModalProcessor[OvisProcessingInfo]):
     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 @@ class Ovis2_5MultiModalProcessor(BaseMultiModalProcessor[Ovis2_5ProcessingInfo])
     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 @@ class SiglipVisionModel(nn.Module):
     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.",