GCG推理时无法批处理

作者您好，我发现在我移植的gcg推理代码中无法实现批处理，具体如下：
predict, seg_hidden_states = forward_model(
batch_inputs, pixel_values,
self.omg_tokenizer, self.omg_model, self.omg_llm,
self.omg_projector_text2vision,
self.omg_gen_config, self.omg_stop_criteria)
对于batch为4的图像输入，pixel_values包含四个元素，如下：

但是predict, seg_hidden_statesg尺寸不对，是对应单图的尺寸，如下：

怀疑是调用代码的这一句有问题，当pixel_values为单个或多个元素的list时，mm['inputs_embeds']都为(1,68,4096)的向量
mm_inputs = prepare_inputs_labels_for_multimodal(llm=llm, input_ids=ids, pixel_values=pixel_values)

Start with the GCG inference call to forward_model and the named prepare_inputs_labels_for_multimodal function. Reproduce the issue using four image inputs and inspect mm['inputs_embeds'], predict, and seg_hidden_states. Done means batched inputs produce outputs with the corresponding batch dimensions rather than single-image shapes.