[Bugfix] Fix embedding assignment for InternVL-based models (#15086)

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

examples/offline_inference/vision_language.py

@@ -169,7 +169,6 @@ def run_gemma3(questions: list[str], modality: str) -> ModelRequestData:
         model=model_name,
         max_model_len=2048,
         max_num_seqs=2,
-        # Default is False; setting it to True is not supported in V1 yet
         mm_processor_kwargs={"do_pan_and_scan": True},
         disable_mm_preprocessor_cache=args.disable_mm_preprocessor_cache,
     )

examples/offline_inference/vision_language_multi_image.py

@@ -91,8 +91,6 @@ def load_gemma3(question: str, image_urls: list[str]) -> ModelRequestData:
         model=model_name,
         max_model_len=8192,
         max_num_seqs=2,
-        # Default is False; setting it to True is not supported in V1 yet
-        mm_processor_kwargs={"do_pan_and_scan": True},
         limit_mm_per_prompt={"image": len(image_urls)},
     )
 

vllm/model_executor/models/gemma3_mm.py

@@ -183,7 +183,7 @@ class Gemma3ProcessingInfo(BaseProcessingInfo):
         image_width: int,
         image_height: int,
         processor: Optional[Gemma3Processor],
-    ) -> PromptUpdateDetails:
+    ) -> PromptUpdateDetails[str]:
         if processor is None:
             processor = self.get_hf_processor()
 

vllm/model_executor/models/h2ovl.py

@@ -249,20 +249,15 @@ class H2OVLProcessor(BaseInternVLProcessor):
     def image_token_id(self) -> int:
         return self.tokenizer.get_vocab()[IMG_CONTEXT]
 
-    def get_image_repl_features(
+    def get_image_repl(
         self,
         feature_size: int,
         num_patches: Optional[int],
-    ) -> str:
-        return IMG_CONTEXT * feature_size
+    ) -> PromptUpdateDetails[str]:
+        repl_features = IMG_CONTEXT * feature_size
+        repl_full = IMG_START + repl_features + IMG_END
 
-    def get_image_repl_full(
-        self,
-        feature_size: int,
-        num_patches: Optional[int],
-    ) -> str:
-        features = self.get_image_repl_features(feature_size, num_patches)
-        return IMG_START + features + IMG_END
+        return PromptUpdateDetails(full=repl_full, features=repl_features)
 
     def resolve_min_max_num(
         self,
@@ -501,12 +496,7 @@ class H2OVLMultiModalProcessor(InternVLMultiModalProcessor[H2OVLProcessingInfo]
             if num_patches is not None:
                 assert isinstance(num_patches, int)
 
-            return PromptUpdateDetails(
-                full=hf_processor.get_image_repl_full(feature_size,
-                                                      num_patches),
-                features=hf_processor.get_image_repl_features(
-                    feature_size, num_patches),
-            )
+            return hf_processor.get_image_repl(feature_size, num_patches)
 
         return [
             PromptReplacement(

vllm/model_executor/models/internvl.py

@@ -9,14 +9,13 @@
 from abc import ABC, abstractmethod
 from collections.abc import Iterable, Mapping, Sequence
 from functools import cached_property
-from typing import (List, Literal, Optional, Set, Tuple, TypedDict, TypeVar,
-                    Union)
+from typing import Literal, Optional, Set, Tuple, TypedDict, TypeVar, Union
 
 import torch
 import torch.nn as nn
 import torchvision.transforms as T
 from PIL import Image
-from transformers import BatchFeature, PretrainedConfig, TensorType
+from transformers import BatchEncoding, PretrainedConfig, TensorType
 
 from vllm.config import VllmConfig
 from vllm.model_executor.layers.quantization import QuantizationConfig
@@ -36,10 +35,12 @@ from vllm.multimodal.processing import (BaseMultiModalProcessor,
 from vllm.multimodal.profiling import BaseDummyInputsBuilder, ProcessorInputs
 from vllm.sequence import IntermediateTensors
 from vllm.transformers_utils.tokenizer import AnyTokenizer
+from vllm.utils import flatten_2d_lists
 
 from .interfaces import MultiModalEmbeddings, SupportsMultiModal, SupportsPP
 from .utils import (AutoWeightsLoader, flatten_bn, init_vllm_registered_model,
                     maybe_prefix, merge_multimodal_embeddings)
+from .vision import scatter_patch_features, select_patch_features
 
 IMG_START = '<img>'
 IMG_END = '</img>'
@@ -51,16 +52,26 @@ IMAGENET_STD = (0.229, 0.224, 0.225)
 
 class InternVLImagePixelInputs(TypedDict):
     type: Literal["pixel_values"]
-    data: torch.Tensor
+    pixel_values_flat: torch.Tensor
     """
     Shape:
     `(batch_size * num_images * (1 + num_patches), num_channels, height, width)`
     """
-    patches_per_image: List[int]
+
+    num_patches: torch.Tensor
+    """Shape: `(batch_size * num_images)`"""
+
+    embed_is_patch: Union[torch.Tensor, list[torch.Tensor]]
     """
-    List of number of total patches for each image in the batch.
+    A boolean mask indicating which image embeddings correspond
+    to patch tokens.
+
+    Shape: `(batch_size, num_images, num_embeds)`
     """
 
+    num_embeds: Union[torch.Tensor, list[torch.Tensor]]
+    """Shape: `(batch_size, num_images)`"""
+
 
 class InternVLImageEmbeddingInputs(TypedDict):
     type: Literal["image_embeds"]
@@ -286,19 +297,11 @@ class BaseInternVLProcessor(ABC):
         raise NotImplementedError
 
     @abstractmethod
-    def get_image_repl_features(
+    def get_image_repl(
         self,
         feature_size: int,
         num_patches: Optional[int],
-    ) -> str:
-        raise NotImplementedError
-
-    @abstractmethod
-    def get_image_repl_full(
-        self,
-        feature_size: int,
-        num_patches: Optional[int],
-    ) -> str:
+    ) -> PromptUpdateDetails[str]:
         raise NotImplementedError
 
     def resolve_min_max_num(
@@ -394,7 +397,7 @@ class BaseInternVLProcessor(ABC):
         max_dynamic_patch: Optional[int] = None,
         dynamic_image_size: Optional[bool] = None,
         return_tensors: Optional[Union[str, TensorType]] = None,
-    ) -> BatchFeature:
+    ) -> Mapping[str, NestedTensors]:
         if text is None:
             text = []
         if not isinstance(text, list):
@@ -413,28 +416,41 @@ class BaseInternVLProcessor(ABC):
                 max_dynamic_patch=max_dynamic_patch,
                 dynamic_image_size=dynamic_image_size,
             )
-            image_inputs = {
-                "pixel_values_flat": torch.cat(pixel_values_lst),
-                "image_num_patches": list(map(len, pixel_values_lst)),
+            image_inputs: dict[str, NestedTensors] = {
+                "pixel_values_flat":
+                torch.cat(pixel_values_lst),
+                "image_num_patches":
+                torch.tensor([len(item) for item in pixel_values_lst]),
             }
 
+            tokenizer = self.tokenizer
+            image_token_id = self.image_token_id
+
+            num_embeds = list[int]()
+            embed_is_patch = list[torch.Tensor]()
+
             for pixel_values in pixel_values_lst:
                 num_patches = pixel_values.shape[0]
                 feature_size = num_patches * self.num_image_token
 
-                image_repl = self.get_image_repl_full(feature_size,
-                                                      num_patches)
-                text = [t.replace('<image>', image_repl, 1) for t in text]
+                image_repl = self.get_image_repl(feature_size, num_patches)
+                feature_tokens = tokenizer.encode(image_repl.features,
+                                                  add_special_tokens=False)
+
+                text = [t.replace('<image>', image_repl.full, 1) for t in text]
+                num_embeds.append(len(feature_tokens))
+                embed_is_patch.append(
+                    torch.tensor(feature_tokens) == image_token_id)
+
+            image_inputs["num_embeds"] = torch.tensor(num_embeds)
+            image_inputs["embed_is_patch"] = embed_is_patch
 
         text_inputs = self.tokenizer(text)
 
-        return BatchFeature(
-            {
-                **text_inputs,
-                **image_inputs,
-            },
-            tensor_type=return_tensors,
-        )
+        return {
+            **BatchEncoding(text_inputs, tensor_type=return_tensors),
+            **image_inputs,
+        }
 
 
 class InternVLProcessor(BaseInternVLProcessor):
@@ -443,20 +459,15 @@ class InternVLProcessor(BaseInternVLProcessor):
     def image_token_id(self) -> int:
         return self.tokenizer.get_vocab()[IMG_CONTEXT]
 
-    def get_image_repl_features(
+    def get_image_repl(
         self,
         feature_size: int,
         num_patches: Optional[int],
-    ) -> str:
-        return IMG_CONTEXT * feature_size
+    ) -> PromptUpdateDetails[str]:
+        repl_features = IMG_CONTEXT * feature_size
+        repl_full = IMG_START + repl_features + IMG_END
 
-    def get_image_repl_full(
-        self,
-        feature_size: int,
-        num_patches: Optional[int],
-    ) -> str:
-        features = self.get_image_repl_features(feature_size, num_patches)
-        return IMG_START + features + IMG_END
+        return PromptUpdateDetails(full=repl_full, features=repl_features)
 
 
 class BaseInternVLProcessingInfo(BaseProcessingInfo):
@@ -566,16 +577,15 @@ class InternVLMultiModalProcessor(BaseMultiModalProcessor[_I]):
         prompt: str,
         mm_data: Mapping[str, object],
         mm_kwargs: Mapping[str, object],
-    ) -> BatchFeature:
+    ) -> Mapping[str, NestedTensors]:
         processed_outputs = super()._call_hf_processor(
             prompt=prompt,
             mm_data=mm_data,
             mm_kwargs=mm_kwargs,
         )
 
-        image_token_id = self.info.get_hf_processor(**mm_kwargs).image_token_id
-        image_data = mm_data.get("images", [])
-        assert isinstance(image_data, list)
+        hf_processor = self.info.get_hf_processor(**mm_kwargs)
+        image_token_id = hf_processor.image_token_id
 
         # Since there may be extra tokens in the feature placeholders,
         # we need to pass the image token ID to the model to select the
@@ -586,7 +596,7 @@ class InternVLMultiModalProcessor(BaseMultiModalProcessor[_I]):
 
     def _get_mm_fields_config(
         self,
-        hf_inputs: BatchFeature,
+        hf_inputs: Mapping[str, NestedTensors],
         hf_processor_mm_kwargs: Mapping[str, object],
     ) -> Mapping[str, MultiModalFieldConfig]:
         image_num_patches = hf_inputs.get("image_num_patches", torch.empty(0))
@@ -596,6 +606,8 @@ class InternVLMultiModalProcessor(BaseMultiModalProcessor[_I]):
             pixel_values_flat=MultiModalFieldConfig.flat_from_sizes(
                 "image", image_num_patches),
             image_num_patches=MultiModalFieldConfig.batched("image"),
+            embed_is_patch=MultiModalFieldConfig.batched("image"),
+            num_embeds=MultiModalFieldConfig.batched("image"),
             image_embeds=MultiModalFieldConfig.batched("image"),
             image_token_id=MultiModalFieldConfig.shared("image", num_images),
         )
@@ -637,12 +649,7 @@ class InternVLMultiModalProcessor(BaseMultiModalProcessor[_I]):
             if num_patches is not None:
                 assert isinstance(num_patches, int)
 
-            return PromptUpdateDetails(
-                full=hf_processor.get_image_repl_full(feature_size,
-                                                      num_patches),
-                features=hf_processor.get_image_repl_features(
-                    feature_size, num_patches),
-            )
+            return hf_processor.get_image_repl(feature_size, num_patches)
 
         return [
             PromptReplacement(
@@ -832,6 +839,8 @@ class InternVLChatModel(nn.Module, SupportsMultiModal, SupportsPP):
             self, **kwargs: object) -> Optional[InternVLImageInputs]:
         pixel_values_flat = kwargs.pop("pixel_values_flat", None)
         image_num_patches = kwargs.pop("image_num_patches", None)
+        embed_is_patch = kwargs.pop("embed_is_patch", None)
+        num_embeds = kwargs.pop("num_embeds", None)
         image_embeds = kwargs.pop("image_embeds", None)
 
         if pixel_values_flat is None and image_embeds is None:
@@ -858,35 +867,47 @@ class InternVLChatModel(nn.Module, SupportsMultiModal, SupportsPP):
 
             if not isinstance(image_num_patches, (torch.Tensor, list)):
                 raise ValueError("Incorrect type of image_num_patches. "
-                                 f"Got type: {type(pixel_values_flat)}")
+                                 f"Got type: {type(image_num_patches)}")
+
+            if not isinstance(embed_is_patch, (torch.Tensor, list)):
+                raise ValueError("Incorrect type of embed_is_patch. "
+                                 f"Got type: {type(embed_is_patch)}")
+
+            if not isinstance(num_embeds, (torch.Tensor, list)):
+                raise ValueError("Incorrect type of num_embeds. "
+                                 f"Got type: {type(num_embeds)}")
+
+            pixel_values_flat = flatten_bn(pixel_values_flat, concat=True)
+            image_num_patches = flatten_bn(image_num_patches, concat=True)
 
             return InternVLImagePixelInputs(
                 type="pixel_values",
-                data=self._validate_pixel_values(
-                    flatten_bn(pixel_values_flat, concat=True)),
-                patches_per_image=flatten_bn(image_num_patches,
-                                             concat=True).tolist())
+                pixel_values_flat=self._validate_pixel_values(
+                    pixel_values_flat),
+                num_patches=image_num_patches,
+                embed_is_patch=embed_is_patch,
+                num_embeds=num_embeds,
+            )
 
         raise AssertionError("This line should be unreachable.")
 
     def _process_image_input(
         self,
         image_input: InternVLImageInputs,
-    ) -> tuple[torch.Tensor, ...]:
+    ) -> Union[torch.Tensor, tuple[torch.Tensor, ...]]:
         if image_input["type"] == "image_embeds":
             return image_input["data"]
 
         assert self.vision_model is not None
 
-        image_embeds = self.extract_feature(image_input["data"])
+        image_embeds = self.extract_feature(image_input["pixel_values_flat"])
 
-        patches_per_image = image_input["patches_per_image"]
+        num_patches = image_input["num_patches"]
 
         # Only one image in the current batch
-        if len(patches_per_image) == 1:
-            image_embeds = image_embeds.view(
+        if len(num_patches) == 1:
+            return image_embeds.view(
                 -1, self.config.text_config.hidden_size).unsqueeze(0)
-            return image_embeds
 
         # NOTE: Image embeddings are split into separate tensors for each image
         # by the size of each embedding.
@@ -894,10 +915,9 @@ class InternVLChatModel(nn.Module, SupportsMultiModal, SupportsPP):
         image_embeds = image_embeds.view(-1,
                                          self.config.text_config.hidden_size)
         image_feature_sizes = [
-            num_patches * feature_size for num_patches in patches_per_image
+            num_patches * feature_size for num_patches in num_patches
         ]
-        image_embeds = image_embeds.split(image_feature_sizes)
-        return image_embeds
+        return image_embeds.split(image_feature_sizes)
 
     def _set_visual_token_mask(self, input_ids: torch.Tensor) -> None:
         if self.is_mono:
@@ -911,8 +931,19 @@ class InternVLChatModel(nn.Module, SupportsMultiModal, SupportsPP):
         image_input = self._parse_and_validate_image_input(**kwargs)
         if image_input is None:
             return None
-        vision_embeddings = self._process_image_input(image_input)
-        return vision_embeddings
+
+        image_features = self._process_image_input(image_input)
+
+        if (kwargs.get("v0_path", False)
+                or image_input["type"] != "pixel_values"):
+            return image_features
+
+        return flatten_2d_lists(
+            scatter_patch_features(*args) for args in zip(
+                image_features,
+                image_input["num_embeds"],
+                image_input["embed_is_patch"],
+            ))
 
     def get_input_embeddings(
         self,
@@ -924,8 +955,11 @@ class InternVLChatModel(nn.Module, SupportsMultiModal, SupportsPP):
             assert self.img_context_token_id is not None
             self._set_visual_token_mask(input_ids)
             inputs_embeds = merge_multimodal_embeddings(
-                input_ids, inputs_embeds, multimodal_embeddings,
-                self.img_context_token_id)
+                input_ids,
+                inputs_embeds,
+                select_patch_features(multimodal_embeddings),
+                self.img_context_token_id,
+            )
         return inputs_embeds
 
     def forward(
@@ -944,6 +978,7 @@ class InternVLChatModel(nn.Module, SupportsMultiModal, SupportsPP):
         # NOTE: In v1, inputs_embeds is always generated at model runner, this
         # condition is for v0 compatibility.
         elif inputs_embeds is None:
+            kwargs.update({"v0_path": True})
             vision_embeddings = self.get_multimodal_embeddings(**kwargs)
             inputs_embeds = self.get_input_embeddings(input_ids,
                                                       vision_embeddings)

vllm/model_executor/models/nvlm_d.py

@@ -36,11 +36,11 @@ class NVLMProcessor(BaseInternVLProcessor):
     def image_token_id(self) -> int:
         return self.tokenizer.get_vocab()[IMG_PAD]
 
-    def get_image_repl_features(
+    def get_image_repl(
         self,
         feature_size: int,
         num_patches: Optional[int],
-    ) -> str:
+    ) -> PromptUpdateDetails[str]:
         if num_patches is None:
             raise NotImplementedError("Embedding inputs are not supported")
 
@@ -55,14 +55,9 @@ class NVLMProcessor(BaseInternVLProcessor):
         # We include the start and end as well because "<Image><tile" is
         # tokenized as ["<Image", "><", "tile"], resulting in assertion error
         # when trying to find "<tile" as a subsequence of "<Image><tile"
-        return "<Image>" + features + "</Image>"
+        repl = "<Image>" + features + "</Image>"
 
-    def get_image_repl_full(
-        self,
-        feature_size: int,
-        num_patches: Optional[int],
-    ) -> str:
-        return self.get_image_repl_features(feature_size, num_patches)
+        return PromptUpdateDetails(full=repl, features=repl)
 
 
 class NVLMProcessingInfo(BaseInternVLProcessingInfo):
@@ -180,11 +175,11 @@ class NVLMMultiModalProcessor(InternVLMultiModalProcessor[NVLMProcessingInfo]):
             if num_patches is not None:
                 assert isinstance(num_patches, int)
 
+            repl = hf_processor.get_image_repl(feature_size, num_patches)
+
             return PromptUpdateDetails(
-                full=hf_processor.get_image_repl_full(feature_size,
-                                                      num_patches) + "\n",
-                features=hf_processor.get_image_repl_features(
-                    feature_size, num_patches) + "\n",
+                full=repl.full + "\n",
+                features=repl.features + "\n",
             )
 
         # See note in dummy data regarding why we have the extra newline

vllm/multimodal/processing.py

@@ -103,13 +103,13 @@ The token sequence or text to update.
 
 
 @dataclass
-class PromptUpdateDetails:
+class PromptUpdateDetails(Generic[_S]):
     """Details about the token sequence or text that are part of the update."""
 
-    full: PromptSeq
+    full: _S
     """The full content."""
 
-    features: PromptSeq
+    features: _S
     """
     The part of the content that corresponds to feature placeholders;
     this will be replaced by the output of the vision encoder during model
@@ -117,7 +117,7 @@ class PromptUpdateDetails:
     """
 
     @staticmethod
-    def from_seq(seq: PromptSeq) -> "PromptUpdateDetails":
+    def from_seq(seq: _S) -> "PromptUpdateDetails[_S]":
         return PromptUpdateDetails(full=seq, features=seq)