Support mm_processor_cache in the Transformers multimodal backend

docs/features/multimodal_inputs.md

@@ -477,6 +477,11 @@ You must enable this feature via `enable_mm_embeds=True`.
     The vLLM engine may crash if incorrect shape of embeddings is passed.
     Only enable this flag for trusted users!
 
+!!! note
+    The [Transformers modeling backend](../models/supported_models.md#transformers) does not support this feature.
+    Passing pre-computed embeddings skips the Hugging Face processor, which is what provides the extra fields
+    the backend needs to locate the placeholders in the prompt.
+
 #### Image Embeddings
 
 ??? code

tests/models/multimodal/processing/test_transformers_image.py

@@ -5,6 +5,7 @@
 from vllm.assets.image import ImageAsset
 from vllm.config import ModelConfig
 from vllm.multimodal import MULTIMODAL_REGISTRY
+from vllm.multimodal.cache import MultiModalProcessorOnlyCache
 
 
 @pytest.mark.parametrize("model_id", ["llava-hf/llava-onevision-qwen2-0.5b-ov-hf"])
@@ -77,3 +78,25 @@ def test_image_multiple_inputs():
 
     assert len(result["mm_placeholders"]["image"]) == 2
     assert len(result["mm_kwargs"]["image"]) == 2
+
+
+@pytest.mark.parametrize(
+    "model_id",
+    ["google/gemma-3-4b-it", "llava-hf/llava-onevision-qwen2-0.5b-ov-hf"],
+)
+def test_image_cached_apply(model_id):
+    model_config = ModelConfig(model=model_id, model_impl="transformers")
+    cache = MultiModalProcessorOnlyCache(model_config)
+    mm_processor = MULTIMODAL_REGISTRY.create_processor(model_config, cache=cache)
+
+    image = ImageAsset("cherry_blossom").pil_image
+    image_token = mm_processor.info.get_hf_processor().image_token
+    prompt = f"{image_token} What is the content of this image?"
+
+    for _ in range(2):
+        result = mm_processor(
+            prompt=prompt,
+            mm_items=mm_processor.info.parse_mm_data({"image": image}),
+            hf_processor_mm_kwargs={},
+        )
+        assert len(result["mm_placeholders"]["image"]) == 1