[Bugfix] Merge MM embeddings by index instead of token IDs

docs/contributing/model/multimodal.md

@@ -66,35 +66,12 @@ Further update the model as follows:
 !!! important
     The returned `multimodal_embeddings` must be either a **3D [torch.Tensor][]** of shape `(num_items, feature_size, hidden_size)`, or a **list / tuple of 2D [torch.Tensor][]'s** of shape `(feature_size, hidden_size)`, so that `multimodal_embeddings[i]` retrieves the embeddings generated from the `i`-th multimodal data item (e.g, image) of the request.
 
-- Implement [get_input_embeddings][vllm.model_executor.models.interfaces.SupportsMultiModal.get_input_embeddings] to merge `multimodal_embeddings` with text embeddings from the `input_ids`. If input processing for the model is implemented correctly (see sections below), then you can leverage the utility function we provide to easily merge the embeddings.
-
-    ??? code
-
-        ```python
-        from .utils import merge_multimodal_embeddings
-
-        class YourModelForImage2Seq(nn.Module):
-            ...
-
-            def get_input_embeddings(
-                self,
-                input_ids: torch.Tensor,
-                multimodal_embeddings: Optional[MultiModalEmbeddings] = None,
-            ) -> torch.Tensor:
-
-                # `get_input_embeddings` should already be implemented for the language 
-                # model as one of the requirements of basic vLLM model implementation.
-                inputs_embeds = self.language_model.get_input_embeddings(input_ids)
-
-                if multimodal_embeddings is not None:
-                    inputs_embeds = merge_multimodal_embeddings(
-                        input_ids=input_ids, 
-                        inputs_embeds=inputs_embeds, 
-                        multimodal_embeddings=multimodal_embeddings,
-                        placeholder_token_id=self.config.image_token_index)
+!!! note
+    By default, vLLM merges the multimodal embeddings into text embeddings depending on the information of their locations defined in
+    [PlaceholderRange][vllm.multimodal.inputs.PlaceholderRange] from input processing.
+    This logic can be found at [get_input_embeddings][vllm.model_executor.models.interfaces.SupportsMultiModal.get_input_embeddings].
 
-                return inputs_embeds
-        ```
+    You may override this method if additional logic is required for your model when merging embeddings. 
 
 - Implement [get_language_model][vllm.model_executor.models.interfaces.SupportsMultiModal.get_language_model] getter to provide stable access to the underlying language model.
 

vllm/config/model.py

@@ -509,9 +509,14 @@ def _task_to_convert(task: TaskOption) -> ConvertType:
                 else:  # task == "auto"
                     pass
             else:
+                debug_info = {
+                    "architectures": architectures,
+                    "is_generative_model": is_generative_model,
+                    "is_pooling_model": is_pooling_model,
+                }
                 raise AssertionError("The model should be a generative or "
                                      "pooling model when task is set to "
-                                     f"{self.task!r}.")
+                                     f"{self.task!r}. Found: {debug_info}")
 
             self.runner = runner
             self.convert = convert

vllm/model_executor/models/aria.py

@@ -38,8 +38,7 @@
 from .interfaces import MultiModalEmbeddings, SupportsMultiModal, SupportsQuant
 from .llama import LlamaDecoderLayer, LlamaMLP, LlamaModel
 from .utils import (AutoWeightsLoader, WeightsMapper, flatten_bn,
-                    is_pp_missing_parameter, maybe_prefix,
-                    merge_multimodal_embeddings)
+                    is_pp_missing_parameter, maybe_prefix)
 
 
 class AriaImagePixelInputs(TensorSchema):
@@ -605,19 +604,6 @@ def get_multimodal_embeddings(self,
         multimodal_embeddings = self._process_image_input(image_input)
         return multimodal_embeddings
 
-    def get_input_embeddings(
-        self,
-        input_ids: torch.Tensor,
-        multimodal_embeddings: Optional[MultiModalEmbeddings] = None,
-    ) -> torch.Tensor:
-        inputs_embeds = self.language_model.get_input_embeddings(input_ids)
-        if multimodal_embeddings is not None \
-            and len(multimodal_embeddings) != 0:
-            inputs_embeds = merge_multimodal_embeddings(
-                input_ids, inputs_embeds, multimodal_embeddings,
-                self.config.image_token_index)
-        return inputs_embeds
-
     def forward(
         self,
         input_ids: torch.Tensor,
@@ -628,10 +614,11 @@ def forward(
     ) -> Union[torch.Tensor, IntermediateTensors]:
         if inputs_embeds is None:
             multimodal_embeddings = self.get_multimodal_embeddings(**kwargs)
-            # always pass the input via `inputs_embeds`
-            # to make sure the computation graph is consistent
-            inputs_embeds = self.get_input_embeddings(input_ids,
-                                                      multimodal_embeddings)
+            inputs_embeds = self.get_input_embeddings(
+                input_ids,
+                multimodal_embeddings,
+                is_multimodal=input_ids == self.config.image_token_index,
+            )
             input_ids = None
 
         hidden_states = self.language_model(

vllm/model_executor/models/aya_vision.py

@@ -33,8 +33,7 @@
 from .interfaces import MultiModalEmbeddings, SupportsMultiModal, SupportsPP
 from .siglip import SiglipVisionModel
 from .utils import (AutoWeightsLoader, WeightsMapper, flatten_bn,
-                    init_vllm_registered_model, maybe_prefix,
-                    merge_multimodal_embeddings)
+                    init_vllm_registered_model, maybe_prefix)
 
 
 class AyaVisionImagePixelInputs(TensorSchema):
@@ -417,23 +416,6 @@ def get_multimodal_embeddings(self,
 
         return self._process_image_input(image_input, **kwargs)
 
-    def get_input_embeddings(
-        self,
-        input_ids: torch.Tensor,
-        multimodal_embeddings: Optional[MultiModalEmbeddings] = None,
-    ) -> torch.Tensor:
-        inputs_embeds = self.language_model.get_input_embeddings(input_ids)
-        if multimodal_embeddings is not None \
-            and len(multimodal_embeddings) != 0:
-            inputs_embeds = merge_multimodal_embeddings(
-                input_ids=input_ids,
-                inputs_embeds=inputs_embeds,
-                multimodal_embeddings=multimodal_embeddings,
-                placeholder_token_id=self.config.image_token_index,
-            )
-
-        return inputs_embeds
-
     def forward(
         self,
         input_ids: torch.Tensor,
@@ -449,8 +431,11 @@ def forward(
         # condition is for v0 compatibility.
         elif inputs_embeds is None:
             vision_embeddings = self.get_multimodal_embeddings(**kwargs)
-            inputs_embeds = self.get_input_embeddings(input_ids,
-                                                      vision_embeddings)
+            inputs_embeds = self.get_input_embeddings(
+                input_ids,
+                vision_embeddings,
+                is_multimodal=input_ids == self.config.image_token_index,
+            )
             input_ids = None
 
         hidden_states = self.language_model.model(

vllm/model_executor/models/bert.py

@@ -348,6 +348,9 @@ def __init__(
         self.encoder = BertEncoder(vllm_config=vllm_config,
                                    prefix=f"{prefix}.encoder")
 
+    def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
+        return self.embeddings(input_ids)
+
     def forward(
         self,
         input_ids: torch.Tensor,
@@ -457,6 +460,9 @@ def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
                                        prefix=maybe_prefix(prefix, "model"))
         self.pooler = self._build_pooler(pooler_config)
 
+    def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
+        return self.model.get_input_embeddings(input_ids)
+
     def forward(
         self,
         input_ids: torch.Tensor,
@@ -588,6 +594,9 @@ def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
             ),
         })
 
+    def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
+        return self.bert.get_input_embeddings(input_ids)
+
     def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]):
         loader = AutoWeightsLoader(self)
         loaded_params = loader.load_weights(weights)
@@ -637,6 +646,9 @@ def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
             Pooler.for_encode(pooler_config),
         })
 
+    def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
+        return self.bert.get_input_embeddings(input_ids)
+
     def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]):
         loader = AutoWeightsLoader(self)
         loaded_params = loader.load_weights(weights)

vllm/model_executor/models/bert_with_rope.py

@@ -426,6 +426,9 @@ def __init__(self,
             prefix=f"{prefix}.encoder")
         self.pooler = BertPooler(self.config) if add_pooling_layer else None
 
+    def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
+        return self.embeddings(input_ids)
+
     def forward(
         self,
         input_ids: torch.Tensor,
@@ -673,6 +676,9 @@ def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]):
         loaded_params = loader.load_weights(weights)
         return loaded_params
 
+    def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
+        return self.new.get_input_embeddings(input_ids)
+
     def forward(
         self,
         input_ids: Optional[torch.Tensor],

vllm/model_executor/models/blip2.py

@@ -27,7 +27,7 @@
 from .interfaces import (MultiModalEmbeddings, SupportsMultiModal, SupportsPP,
                          SupportsQuant)
 from .utils import (AutoWeightsLoader, flatten_bn, init_vllm_registered_model,
-                    maybe_prefix, merge_multimodal_embeddings)
+                    maybe_prefix)
 
 # We use this internally as placeholders since there is no image token
 # defined on the HuggingFace repo
@@ -631,19 +631,6 @@ def get_multimodal_embeddings(self,
         vision_embeddings = self._process_image_input(image_input)
         return vision_embeddings
 
-    def get_input_embeddings(
-        self,
-        input_ids: torch.Tensor,
-        multimodal_embeddings: Optional[MultiModalEmbeddings] = None,
-    ) -> torch.Tensor:
-        inputs_embeds = self.language_model.get_input_embeddings(input_ids)
-        if multimodal_embeddings is not None \
-            and len(multimodal_embeddings) != 0:
-            inputs_embeds = merge_multimodal_embeddings(
-                input_ids, inputs_embeds, multimodal_embeddings,
-                _IMAGE_TOKEN_ID)
-        return inputs_embeds
-
     def forward(
         self,
         input_ids: torch.Tensor,
@@ -689,8 +676,11 @@ def forward(
         # condition is for v0 compatibility.
         elif inputs_embeds is None:
             vision_embeddings = self.get_multimodal_embeddings(**kwargs)
-            inputs_embeds = self.get_input_embeddings(input_ids,
-                                                      vision_embeddings)
+            inputs_embeds = self.get_input_embeddings(
+                input_ids,
+                vision_embeddings,
+                is_multimodal=input_ids == _IMAGE_TOKEN_ID,
+            )
             input_ids = None
 
         hidden_states = self.language_model.model(input_ids,

vllm/model_executor/models/chameleon.py

@@ -44,7 +44,7 @@
                          SupportsQuant)
 from .utils import (flatten_bn, is_pp_missing_parameter,
                     make_empty_intermediate_tensors_factory, make_layers,
-                    maybe_prefix, merge_multimodal_embeddings)
+                    maybe_prefix)
 
 logger = init_logger(__name__)
 
@@ -1002,20 +1002,6 @@ def get_multimodal_embeddings(self,
         vision_embeddings = self.model.get_input_embeddings(image_tokens)
         return vision_embeddings
 
-    def get_input_embeddings(
-        self,
-        input_ids: torch.Tensor,
-        multimodal_embeddings: Optional[MultiModalEmbeddings] = None,
-    ) -> torch.Tensor:
-
-        inputs_embeds = self.model.get_input_embeddings(input_ids)
-        if multimodal_embeddings is not None \
-            and len(multimodal_embeddings) != 0:
-            inputs_embeds = merge_multimodal_embeddings(
-                input_ids, inputs_embeds, multimodal_embeddings,
-                self.model.vocabulary_mapping.image_token_id)
-        return inputs_embeds
-
     def forward(
         self,
         input_ids: torch.Tensor,
@@ -1032,8 +1018,12 @@ def forward(
         # condition is for v0 compatibility.
         elif inputs_embeds is None:
             vision_embeddings = self.get_multimodal_embeddings(**kwargs)
-            inputs_embeds = self.get_input_embeddings(input_ids,
-                                                      vision_embeddings)
+            image_token_id = self.model.vocabulary_mapping.image_token_id
+            inputs_embeds = self.get_input_embeddings(
+                input_ids,
+                vision_embeddings,
+                is_multimodal=input_ids == image_token_id,
+            )
             input_ids = None
 
         hidden_states = self.model(input_ids,

vllm/model_executor/models/chatglm.py

@@ -433,6 +433,9 @@ def __init__(
         self.make_empty_intermediate_tensors = (
             self.transformer.make_empty_intermediate_tensors)
 
+    def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
+        return self.transformer.get_input_embeddings(input_ids)
+
     def compute_logits(
         self,
         hidden_states: torch.Tensor,

vllm/model_executor/models/cohere2_vision.py

@@ -37,8 +37,7 @@
 from .interfaces import MultiModalEmbeddings, SupportsMultiModal, SupportsPP
 from .siglip import SiglipVisionModel
 from .utils import (AutoWeightsLoader, WeightsMapper, flatten_bn,
-                    init_vllm_registered_model, maybe_prefix,
-                    merge_multimodal_embeddings)
+                    init_vllm_registered_model, maybe_prefix)
 
 
 class Cohere2VisionImagePixelInputs(TensorSchema):
@@ -430,23 +429,6 @@ def get_multimodal_embeddings(self,
 
         return self._process_image_input(image_input, **kwargs)
 
-    def get_input_embeddings(
-        self,
-        input_ids: torch.Tensor,
-        multimodal_embeddings: Optional[MultiModalEmbeddings] = None,
-    ) -> torch.Tensor:
-        inputs_embeds = self.language_model.get_input_embeddings(input_ids)
-        if multimodal_embeddings is not None \
-            and len(multimodal_embeddings) != 0:
-            inputs_embeds = merge_multimodal_embeddings(
-                input_ids=input_ids,
-                inputs_embeds=inputs_embeds,
-                multimodal_embeddings=multimodal_embeddings,
-                placeholder_token_id=self.config.image_token_id,
-            )
-
-        return inputs_embeds
-
     def forward(
         self,
         input_ids: torch.Tensor,
@@ -462,8 +444,11 @@ def forward(
         # condition is for v0 compatibility.
         elif inputs_embeds is None:
             vision_embeddings = self.get_multimodal_embeddings(**kwargs)
-            inputs_embeds = self.get_input_embeddings(input_ids,
-                                                      vision_embeddings)
+            inputs_embeds = self.get_input_embeddings(
+                input_ids,
+                vision_embeddings,
+                is_multimodal=input_ids == self.config.image_token_id,
+            )
             input_ids = None
 
         hidden_states = self.language_model.model(

vllm/model_executor/models/deepseek_eagle.py

@@ -66,6 +66,9 @@ def __init__(
         self.norm = RMSNorm(self.config.hidden_size,
                             eps=self.config.rms_norm_eps)
 
+    def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
+        return self.embed_tokens(input_ids)
+
     def forward(
         self,
         input_ids: torch.Tensor,
@@ -205,6 +208,9 @@ def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
         self.logits_processor = LogitsProcessor(self.config.vocab_size,
                                                 scale=logit_scale)
 
+    def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
+        return self.model.get_input_embeddings(input_ids)
+
     def forward(
         self,
         input_ids: torch.Tensor,

vllm/model_executor/models/deepseek_mtp.py

@@ -101,6 +101,9 @@ def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
         )
         self.logits_processor = LogitsProcessor(config.vocab_size)
 
+    def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
+        return self.embed_tokens(input_ids)
+
     def forward(
         self,
         input_ids: torch.Tensor,
@@ -142,6 +145,9 @@ def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
                                                  prefix=maybe_prefix(
                                                      prefix, "model"))
 
+    def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
+        return self.model.get_input_embeddings(input_ids)
+
     def forward(
         self,
         input_ids: torch.Tensor,