Language Model (excluding embedding layer) cannot take inputs_embeds

I have a Pixtral multimodal model that uses a Mistral language decoder. I want to utilize the optimized ONNX Mistral language decoder instead of the native PyTorch implementation. In the forward process, I receive the input embeddings that capture both the text and image information. I want to take this embedding now and feed it into the ONNX model and perform inference with onnxruntime_genai. However, when I create the model parameters with params = og.GeneratorParams(model), it still creates input_ids which my decoder-only model does not have, leaving me with the following error: RuntimeError: Model input was not found: input_ids. Has there been a solution to this issue or a workaround such that my ONNX Mistral model can take in embeddings directly as an input? Otherwise, if I work with just input_ids, I cannot incorporate the image embeddings into the input.

Start with the onnxruntime_genai path around og.GeneratorParams(model) and the model inputs that produce the input_ids error. Check how ONNX Mistral decoder inputs are defined and whether input embeddings are exposed during onnxruntime_genai inference; done means the decoder accepts text-and-image embeddings without requiring input_ids, or the supported workaround is documented.