[Core] Dynamic image size support for VLMs (#5276)

Signed-off-by: Xiaowei Jiang <xwjiang2010@gmail.com>
Co-authored-by: Xiaowei Jiang <xwjiang2010@gmail.com>
Co-authored-by: ywang96 <ywang@roblox.com>
Co-authored-by: xwjiang2010 <87673679+xwjiang2010@users.noreply.github.com>
Co-authored-by: Roger Wang <136131678+ywang96@users.noreply.github.com>

vllm/model_executor/models/clip.py

@@ -8,10 +8,14 @@ from PIL import Image
 from transformers import CLIPVisionConfig
 from transformers.models.clip.modeling_clip import CLIPAttention
 
+from vllm.config import ModelConfig
+from vllm.inputs import LLMInputs
 from vllm.model_executor.layers.activation import get_act_fn
 from vllm.model_executor.layers.linear import (ColumnParallelLinear,
                                                RowParallelLinear)
 from vllm.model_executor.layers.quantization import QuantizationConfig
+from vllm.multimodal.image import (cached_get_tokenizer,
+                                   repeat_and_pad_image_tokens)
 from vllm.sequence import SequenceData
 
 
@@ -64,6 +68,39 @@ def dummy_image_for_clip(
     return {"image": image}
 
 
+def input_processor_for_clip(
+    model_config: ModelConfig,
+    hf_config: CLIPVisionConfig,
+    llm_inputs: LLMInputs,
+    *,
+    image_token_id: int,
+    image_feature_size_override: Optional[int] = None,
+):
+    multi_modal_data = llm_inputs.get("multi_modal_data")
+    if multi_modal_data is None or "image" not in multi_modal_data:
+        return llm_inputs
+
+    tokenizer = cached_get_tokenizer(model_config.tokenizer)
+
+    if image_feature_size_override is None:
+        image_feature_size = get_clip_image_feature_size(hf_config)
+    else:
+        image_feature_size = image_feature_size_override
+
+    new_prompt, new_token_ids = repeat_and_pad_image_tokens(
+        tokenizer,
+        llm_inputs.get("prompt"),
+        llm_inputs["prompt_token_ids"],
+        image_token_id=image_token_id,
+        repeat_count=image_feature_size,
+    )
+
+    # NOTE: Create a defensive copy of the original inputs
+    return LLMInputs(prompt_token_ids=new_token_ids,
+                     prompt=new_prompt,
+                     multi_modal_data=multi_modal_data)
+
+
 # Adapted from https://github.com/huggingface/transformers/blob/v4.39.0/src/transformers/models/clip/modeling_clip.py#L164 # noqa
 class CLIPVisionEmbeddings(nn.Module):