[Misc] Remove more get_input_embeddings_v0 (vllm-project#25857)

Signed-off-by: DarkLight1337 <[email protected]>

Author: DarkLight1337

vllm/model_executor/models/gemma3n_mm.py

@@ -45,8 +45,7 @@
 from .interfaces import (MultiModalEmbeddings, SupportsMultiModal,
                          SupportsTranscription)
 from .utils import (AutoWeightsLoader, WeightsMapper, flatten_bn,
-                    init_vllm_registered_model, maybe_prefix,
-                    merge_multimodal_embeddings)
+                    init_vllm_registered_model, maybe_prefix)
 
 logger = init_logger(__name__)
 

vllm/model_executor/models/keye.py

@@ -41,7 +41,6 @@
 from vllm.multimodal.profiling import BaseDummyInputsBuilder
 from vllm.platforms import _Backend
 from vllm.sequence import IntermediateTensors
-from vllm.transformers_utils.config import uses_mrope
 from vllm.utils import is_list_of
 from vllm.utils.tensor_schema import TensorSchema, TensorShape
 
@@ -50,7 +49,7 @@
 from .siglip import SiglipMLP
 from .utils import (AutoWeightsLoader, WeightsMapper,
                     init_vllm_registered_model, is_pp_missing_parameter,
-                    maybe_prefix, merge_multimodal_embeddings)
+                    maybe_prefix)
 from .vision import get_vit_attn_backend
 
 logger = init_logger(__name__)
@@ -1450,32 +1449,6 @@ def get_multimodal_embeddings(
                 multimodal_embeddings += video_embeddings
         return multimodal_embeddings
 
-    def get_input_embeddings_v0(
-        self,
-        input_ids: torch.Tensor,
-        image_input: Optional[Any] = None,
-        video_input: Optional[Any] = None,
-    ) -> torch.Tensor:
-        inputs_embeds = self.get_input_embeddings(input_ids)
-        if image_input is not None:
-            image_embeds = self._process_image_input(image_input)
-            inputs_embeds = merge_multimodal_embeddings(
-                input_ids,
-                inputs_embeds,
-                image_embeds,
-                placeholder_token_id=self.config.image_token_id,
-            )
-
-        if video_input is not None:
-            video_embeds = self._process_video_input(video_input)
-            inputs_embeds = merge_multimodal_embeddings(
-                input_ids,
-                inputs_embeds,
-                video_embeds,
-                placeholder_token_id=self.config.video_token_id,
-            )
-        return inputs_embeds
-
     def forward(
         self,
         input_ids: torch.Tensor,
@@ -1500,23 +1473,6 @@ def forward(
         if intermediate_tensors is not None:
             inputs_embeds = None
 
-        elif inputs_embeds is None:
-            image_input = self._parse_and_validate_image_input(**kwargs)
-            video_input = self._parse_and_validate_video_input(**kwargs)
-            if image_input is None and video_input is None:
-                inputs_embeds = None
-            else:
-                if uses_mrope(self.config):
-                    assert positions.ndim == 2 and positions.size(0) == 3, (
-                        "multimodal section rotary embedding requires "
-                        f"(3, seq_len) positions, but got {positions.size()}")
-                inputs_embeds = self.get_input_embeddings_v0(
-                    input_ids,
-                    image_input=image_input,
-                    video_input=video_input,
-                )
-                input_ids = None
-
         hidden_states = self.language_model.model(
             input_ids=input_ids,
             positions=positions,

vllm/model_executor/models/phi4_multimodal.py

@@ -44,13 +44,7 @@
 from .idefics2_vision_model import Idefics2VisionTransformer
 from .interfaces import MultiModalEmbeddings, SupportsLoRA, SupportsMultiModal
 from .utils import (AutoWeightsLoader, WeightsMapper, flatten_bn,
-                    init_vllm_registered_model, maybe_prefix,
-                    merge_multimodal_embeddings)
-
-# <|endoftext10|> (see vocab.json in hf model)
-_IMAGE_PLACEHOLDER_TOKEN_ID = 200010
-# <|endoftext11|>
-_AUDIO_PLACEHOLDER_TOKEN_ID = 200011
+                    init_vllm_registered_model, maybe_prefix)
 
 _AUDIO_MAX_SOUNDFILE_SIZE = 241_000
 
@@ -1371,35 +1365,6 @@ def get_multimodal_embeddings(self,
 
         return multimodal_embeddings
 
-    def get_input_embeddings_v0(
-        self,
-        input_ids: torch.Tensor,
-        image_input: Optional[Phi4MMImagePixelInputs] = None,
-        audio_input: Optional[Phi4MMAudioFeatureInputs] = None,
-    ) -> torch.Tensor:
-        audio_projection_mode = 'speech'
-        inputs_embeds = self.get_input_embeddings(input_ids)
-        if image_input is not None:
-            image_embeds = self._process_image_input(image_input)
-            inputs_embeds = merge_multimodal_embeddings(
-                input_ids,
-                inputs_embeds,
-                image_embeds,
-                placeholder_token_id=_IMAGE_PLACEHOLDER_TOKEN_ID,
-            )
-            audio_projection_mode = 'vision'
-
-        if audio_input is not None:
-            audio_embeds = self._process_audio_input(
-                audio_input, audio_projection_mode=audio_projection_mode)
-            inputs_embeds = merge_multimodal_embeddings(
-                input_ids,
-                inputs_embeds,
-                audio_embeds,
-                placeholder_token_id=_AUDIO_PLACEHOLDER_TOKEN_ID,
-            )
-        return inputs_embeds
-
     def forward(
         self,
         input_ids: torch.Tensor,

vllm/model_executor/models/utils.py

@@ -10,6 +10,7 @@
 import torch.nn as nn
 from torch.func import functional_call
 from transformers import PretrainedConfig
+from typing_extensions import deprecated
 
 import vllm.envs as envs
 from vllm.config import VllmConfig
@@ -439,6 +440,9 @@ def _merge_multimodal_embeddings(
     return inputs_embeds
 
 
+@deprecated("`merge_multimodal_embeddings` has been replaced with "
+            "`SupportsMultiModal.get_input_embeddings` and will be "
+            "removed in v0.12.")
 def merge_multimodal_embeddings(
     input_ids: torch.Tensor,
     inputs_embeds: torch.Tensor,